[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"topic-meta-image-generation":3,"topic-articles-image-generation":14},{"slug":4,"tag_slug":5,"title_zh":6,"title_en":7,"intro_zh":8,"intro_en":9,"id":10,"is_active":11,"created_at":12,"modified_at":13},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1",true,"2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"total":15,"page":16,"size":17,"items":18},67,1,100,[19,47,70,91,114,137,160,182,202,221,242,263,282,301,320,341,360,379,397,416,439,458,477,499,518,537,556,575,595,614,632,650,671,692,711,729,747,765,783,801,819,839,857,875,893,911,929,947,965,982,1000,1018,1036,1054,1072,1090,1108,1126,1144,1162,1180,1198,1216,1233,1250,1268,1287],{"id":20,"title":21,"summary":22,"tags":23,"translations":36,"news_slug":43,"published_at":44,"created_at":45,"image_url":27,"view_count":46},"5a1cb663-bfe6-4c9a-bc4e-e1c8e08f942f","单张4090从零训文生图:论文自报160卡时55美元","西班牙四人团队发布 Asuka-FM: 206M 双流 DiT 在单张 RTX 4090 上从零预训练, 论文自报约 160 GPU 小时, 成本不到 55 美元. 论文实测 GenEval 0.609 (512px, 无基准微调), 用内存 latent 缓存, SPRINT 稀疏与像素迁移拆掉 VAE 吞吐墙.",[24,28,31,34],{"id":25,"name":26,"slug":26,"description":27,"color":27},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":29,"name":30,"slug":30,"description":27,"color":27},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":32,"name":33,"slug":33,"description":27,"color":27},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":35,"name":5,"slug":5,"description":27,"color":27},"c883fd20-1d66-4fb7-9fc7-320fa7f87023",[37],{"id":38,"lang":39,"title":40,"summary":41,"content":42},"a587076f-78d8-4843-a747-8362e25dcb4b","en","Asuka-FM: a T2I DiT Pretrained from Scratch on One RTX 4090","A 206M dual-stream DiT pretrained from scratch on one RTX 4090: about 160 GPU hours, under 55 USD. GenEval 0.609 at 512px, no benchmark tuning.","What can 55 USD buy these days? A few days of frontier text-to-image API credits, or a decent dinner. A four-person academic team at Universidad Carlos III de Madrid in Spain offers a third answer: on the paper's own accounting, that money rents a single RTX 4090 for about 160 GPU hours - enough to pretrain a 206M-parameter text-to-image DiT from scratch, not fine-tune anyone else's weights (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26437).\n\n## Why training a T2I model from scratch is hard\n\nThe paper names the incumbents up front: FLUX.1 at 12B, SD3 at 8B, Qwen-Image at 20B, all pretrained on industrial clusters. An academic group starting from zero hits three walls at once. Attention cost grows quadratically with sequence length. Real images come in every aspect ratio, and bucketing them feeds variable shapes into torch.compile, whose JIT graphs keep breaking and recompiling. The quieter wall is the frozen autoencoder. The paper does the arithmetic: pre-computing 32-channel float16 latents for hundreds of thousands of images takes multiple terabytes of fast storage, beyond consumer NVMe, so latents must be encoded on the fly. On a 4090 the paper measures the frozen VAE encoder at 164.92 ms per step, 62.0% of total step latency, while the DiT forward and backward passes together stay under 70 ms. At 1024^2 resolution, on-the-fly VAE encoding takes 72.6% of wall-clock time.\n\n## Four moves to dismantle the walls\n\nFirst, drop per-block AdaLN modulation; time conditioning becomes a prepended continuous sinusoidal time token, and the backbone sheds 28% of its parameters - the paper reports no loss in generative quality. Second, sequence sparsification: SPRINT is prior work by Park et al. (https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.21986), which this paper extends to open-domain multimodal conditioning. 75%-sparse token sequences pass through 9 heavy middle blocks, attention FLOPs drop by more than half, and measured throughput at 512^2 rises from 64.8 to 108.5 img\u002Fs. Third, continuous aspect-normalized RoPE plus shifted square cropping: tensor shapes become fully static, recompilation disappears, and the model gains zero-shot virtual camera control for free. Fourth, a 10M-parameter Patch Detailer Head under velocity regression transfers the latent weights into pixel space - at inference, VAE decoding accounts for over 70% of latency under few-step regimes, and direct RGB output bypasses that wall entirely.\n\nThe training-side VAE wall gets a different fix: at 256^2, pre-encoded latents for all 625K images amount to about 52 GB, which fits in 128 GB of DDR5 system RAM - disk I\u002FO and autoencoder latency vanish together, throughput reaches 282.6 img\u002Fs at 86.8% measured MFU. At 512^2 the latents swell to about 203 GB and no longer fit, so the pipeline switches to network latent streaming. Across the whole pipeline the system sustains an average 75.6% physical MFU on a Ryzen 9 7950X plus one RTX 4090.\n\n## Scoreboard, and the cold water\n\nPaper-reported results: GenEval 0.588 \u002F DPG-Bench 78.96 after the 68-hour 256px stage, and GenEval 0.609 \u002F DPG 78.51 after the 81-hour 512px stage, with no benchmark-specific fine-tuning - the 256px DPG score actually comes out slightly higher than the 512px one. For context, SD3 Medium (2B, cluster) sits at 0.620 \u002F 84.08, and the academic baseline MiniT2I-L\u002F16 (0.91B, 8x H100 for 27 days) reaches 0.883 \u002F 84.80 - though the paper notes those baselines were fine-tuned on BLIP3o-60K, a prompt set that aligns directly with GenEval syntactic patterns. The claim of a new Pareto-optimal efficiency standard is the authors' own.\n\nNow the cold water. The absolute scores still trail by a wide margin; 625K training images against hundreds of millions is not a fair fight, and the authors themselves concede that larger multi-billion models remain stronger on complex crowd scenes and long-tail concepts. Multi-word text rendering is inconsistent. In pixel space, sampling with 13 or more steps runs slower than latent diffusion due to repeated Patch Detailer evaluation. And one accounting note: the abstract says 160 GPU hours (under 55 USD at the paper's stated 0.33 per GPU-hour), while the training roadmap lists 68h, 81h and 57.7h across three stages, summing to about 207 hours. Treat the 55 USD figure as the paper's own headline accounting.\n\n## Not an isolated case, a route\n\nHDM, a single-author project by Shih-Ying Yeh built around a U-shaped Cross-U-Transformer, trained a 1024^2 text-to-image model from scratch on four RTX 5090s for a self-reported 535-620 USD (https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.06068), and the bibliography also cites Sehwag et al.'s CVPR 2025 micro-budget diffusion training. What distinguishes this paper is completeness: a 625K multi-source open corpus plus a 337K curated Danbooru subset, code on GitHub (deeplearningcafe\u002Fasuka-fm) and weights on HuggingFace (aipracticecafe\u002Fasuka-fm) are all released, and the route now extends to the open domain.\n\nSo what. For individuals and small labs, training your own T2I base model has gone from slogan to a 55-USD-scale weekend project, on the paper's accounting. And one outlook line deserves special attention: during RL post-training, generating 12 rollouts per prompt requires 12 independent VAE decodes occupying over 60% of GPU memory in latent mode - direct pixel output makes online RL feasible on consumer workstations. Foundation models no longer inherently belong to datacenters; at least at the 200M scale, that claim now comes with a receipt.","asuka-fm-single-rtx-4090-t2i-pretraining","2026-11-21T13:45:00Z","2026-11-21T13:28:16.722823Z",33,{"id":48,"title":49,"summary":50,"tags":51,"translations":60,"news_slug":66,"published_at":67,"created_at":68,"image_url":27,"view_count":69},"389eb9e6-1232-465f-b278-c073fadae9e8","像素扩散自己长出层级:HiT用四成算力反超JiT","Carnegie Mellon与Rice的HiT给像素空间扩散Transformer加了可学习的token分组层级:相关patch压成组token,组数随图像与扩散步自适应。ImageNet 256拿到1.48 FID,约四成训练算力追平JiT,文生图五基准平均分超过参数量大六倍的UniLiP-3B。",[52,55,56,57],{"id":53,"name":54,"slug":54,"description":27,"color":27},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},{"id":58,"name":59,"slug":59,"description":27,"color":27},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[61],{"id":62,"lang":39,"title":63,"summary":64,"content":65},"fcb2516f-59e6-4385-a068-5c84333e8be9","HiT Learns Its Own Hierarchy, Beats JiT with 40% FLOPs","HiT adds a learned token-grouping hierarchy to pixel-space diffusion: 1.48 FID on ImageNet 256, JiT-level quality at about 40% of its training FLOPs.","The default recipe in image generation is two-stage: train a visual autoencoder separately to squeeze images into a latent space, then run diffusion in that latent space. The end-to-end pixel-space route skips the encoder, and has paid for it in quality and convergence. HiT, the Hierarchical Diffusion Transformer from a Carnegie Mellon and Rice team released on November 17, narrows that gap again, and it does so by growing something new inside the architecture: the model learns by itself which patches may share a token. The paper and code are public (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.22962).\n\n## What it does: let tokens learn to cluster\n\nThe diagnosis in the paper is that the weakness of pixel-space models is not direct pixel modeling but the fact that every patch keeps one token through the whole backbone, so nothing pushes the model to summarize related content at coarser levels. Inspired by H-Net, HiT arranges Transformer blocks into nested encoder, compact, and decoder stages. Outer blocks run at full resolution; a learned similarity router then performs average-linkage clustering, merges similar tokens into groups, and keeps a single token per group. The central trunk processes this much shorter sequence, ungrouping then returns group features to their original pixel positions, with skip connections carrying detail. The token count goes from N to K1 to K2 and back.\n\nNeither memberships nor group counts are fixed. A batch-level compression budget uses a ratio loss to hold average compression near 2x per level, while the actual partition for each image, and at each diffusion timestep, is decided by content. Hard assignments are discrete, so gradients reach the similarity projections through soft pooling and readback weights. The paper stresses this is a budget, not a quota: large backgrounds can collapse into a few groups while complex subjects keep more of them.\n\n## The scorecard: 40% of the FLOPs, passing JiT in the same table\n\nOn ImageNet 256x256, HiT-XL\u002F16 reaches 1.87 FID after 280 epochs, nearly matching JiT-H\u002F16 at 1.86 after 600 epochs with less than half the updates. Per the paper's Figure 1, reaching FID of about 2.16 takes 2.4x less training compute and 2.3x fewer updates; the paper's own framing is comparable quality at roughly 40% of JiT's cumulative training FLOPs. After the full 600 epochs the number is 1.48, ahead of JiT-H's 1.86 in the same table, and above three pixel-space models that lean on DINOv2 external visual supervision (PixNerd 2.15, PixelGen 1.83, PixelDiT 1.61). It approaches REPA-SiT at 1.42 in latent space, though DDT at 1.26 and RAE at 1.13 in that same table are still ahead. Across three matched sizes the ablations all favor HiT: with the same 18 blocks, HiT-B scores 6.61 versus JiT-B's 10.37, the L tier 4.16 versus 4.85, the XL tier 3.64 versus 4.39, each at lower forward FLOPs.\n\nText-to-image (trained on CC12M, at 512x512, with a frozen FLAN-T5-Large text encoder) is where the margins look louder: the 0.25B HiT-B is higher than the same-recipe baseline MiniT2I-B on all five benchmarks, 65.93 average versus 64.64, with 31% fewer forward FLOPs; it also edges past the 1.60B UniLiP-3B at 65.76, with roughly 6x fewer denoiser parameters and 3x fewer forward FLOPs. Worth noting: both baselines, JiT and MiniT2I, come from work Kaiming He is part of.\n\n## No labels, and the hierarchy still emerges\n\nThe most readable part is the emergent behavior. Without any segmentation labels, HiT learns to fold repeated background structures into a few groups, to gather parts of the same object together, and even to place spatially disconnected but visually similar regions, such as car markings, into one group. As the noise level changes across diffusion steps, memberships and group counts reorganize. The text-to-image version lets image tokens and caption words enter one shared hierarchy: in Figure 1, a selected banana patch routes strong weight to the word 'banana', and the group holding 'boy' pools the boy's face and upper body. No word-region annotation exists anywhere in training.\n\n## The cold water\n\nFirst, this is a v1 preprint, listed as Under Review, and every benchmark number is self-reported with no third-party replication yet. Second, the 1.48 headline is not purely architectural credit: it uses group guidance inspired by Sparse Guidance, contrasting a conditional prediction with finer grouping against an unconditional one with coarser grouping at scale 2.6, plus a time-grid shift of k=1.7 in the sampler. Setting the multipliers back to 1 recovers standard CFG, but the paper does not isolate how much the architecture alone contributes. Third, the FLOPs in the tables are nominal costs computed from target group counts; actual cost varies with the input. Fourth, the UniLiP-3B comparison is cross-paper, and how closely training recipes align follows the paper's own statement.\n\n## So what\n\nThe paper's conclusion is stated bluntly: the barrier to stronger image generation was never the lack of external alignment, but the lack of an architecture that learns a visual hierarchy by itself. That claim is single-source for now, but the mechanism lane is not lonely; on the discriminative side, ToMe showed long ago that token merging can double ViT-L@512 throughput with almost no accuracy cost. Whether generative models need a separately pretrained visual encoder is a question worth asking again. The code is out with a quick start (https:\u002F\u002Fgithub.com\u002FNicholas0228\u002FHiT-Released), and model links are on the project page.","hit-pixel-diffusion-learned-hierarchy","2026-11-18T19:50:00Z","2026-11-18T19:13:02.947439Z",82,{"id":71,"title":72,"summary":73,"tags":74,"translations":81,"news_slug":87,"published_at":88,"created_at":89,"image_url":27,"view_count":90},"3f4ca608-fb14-45b2-aeed-122264f141e7","两人团队2.1万美元训出图像模型，账单全公开","两人小团队 Linum 用 8,415 个 H100 小时、21,038 美元训练出 2.18B 参数的像素空间文生图模型 P-JiT，权重与代码开源，并完整公开训练配方：两阶段共 6.6 亿样本、两种优化器分工、噪声调度课程、多分辨率读出方案。官方也自陈这个模型无法在图中写出任何文字。",[75,76,79,80],{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[82],{"id":83,"lang":39,"title":84,"summary":85,"content":86},"035f9f3c-7983-44b1-aae0-3af221879109","Two Brothers Trained a 2B Image Model for $21,038","Linum trained a 2.18B pixel-space model P-JiT for $21,038 on 8,415 H100-hours, releasing the weights, code and the training recipe under Apache 2.0.","There is a habit in image generation of talking about parameter counts and leaderboard scores. On November 4, a two-person lab called Linum published a different kind of number: the bill. Their text-to-image model, Pyramid-JiT, was trained for a total of 8,415 H100-hours, which the lab prices at $21,038. The weights and the code are released under Apache 2.0, and the full training recipe — sample counts, batch size, optimizer split, noise schedule, learning-rate curve — is published alongside them (field notes: https:\u002F\u002Fwww.linum.ai\u002Ffield-notes\u002Fpjit-21k).\n\nThe architecture is the more interesting part of the story, and the reason a two-person team would bother publishing a bill at all.\n\n## Throwing away the VAE\n\nMost text-to-image systems work in latent space: a VAE compresses an image into a lower-dimensional latent, and the diffusion model learns to generate inside that subspace. Linum argues this route hits an empirical ceiling. Their claim is that 16x16 token compression is about as far as latent diffusion models go without surgery on the autoencoder, and that improving a VAE’s reconstruction quality can make the downstream generations worse rather than better.\n\nSo they removed the VAE. The diffusion model operates directly on pixels and learns a latent space purpose-built for generation. The price of that choice is that each token carries more information, so patches grew from 16x16 to 32x32: a 32x32x3 = 3,072-value RGB patch is projected into 1,024 dimensions, a 3:1 compression rate, and the attention context window shrinks by a factor of four relative to a 16x16 setup. Their previous model, Linum v2, was a fairly standard Wan VAE plus DiT with 16x16 token reduction, the same as most models.\n\nThe architecture, Pyramid-JiT, runs 22 DiT blocks in the trunk, preceded by two image refiner blocks and two text refiner blocks, with a hidden size of 2,944 and 23 heads of width 128. Text comes from a frozen Qwen3.5-4B encoder, read at layers 7, 15 and 27. Sampling uses 50 Euler steps with adaptive projected guidance at scale 15, and the released checkpoint is 8.7 GB in fp32 alongside an 8 GB bf16 text encoder.\n\n## Letting the trunk draft its own answer\n\nThe more unusual design choice is the readouts. Because the model is trained with x-prediction in pixel space, every intermediate DiT block is already implicitly predicting a version of the final image. Linum exploits that: a two-layer PixelREPA readout at block 5 aligns the model’s hidden state to DINOv3-L features, and two linear readout heads after blocks 10 and 16 predict the clean image at quarter and half resolution. These heads are training-only and add about 0.24B parameters, bringing the training-time count to roughly 2.42B against 2.18B at sampling time.\n\nIn their October 6 architecture post, the comparison is direct. On the same 22-block trunk with the same recipe for 75M samples, a single readout at 512x512 gives a validation loss of 0.1037. Switching to three readouts — 128x128 at block 10, 256x256 at block 16, 512x512 at block 22 — drops it to 0.0934, a 10.0 percent reduction. Three readouts at full resolution give 0.0948, an 8.5 percent reduction, and down-weighting the low-resolution readouts gives 0.0983, a 5.1 percent reduction. The same post claims P-JiT matches Linum v2’s final FD-DINOv2 score using 11.3 times fewer training samples. That is a comparison against the lab’s own previous model, not against other labs’ models.\n\n## What the bill actually contains\n\nPretraining cost $20,348, which is 8,139 H100-hours priced at $2.50 per hour. Post-training cost $690, or 276 H100-hours, using a variant of Finite Difference Flow Optimization with JEV-Image and human-in-the-loop judges. The hardware was 64 H100s across 8 nodes, in bf16 with FSDP2, 32 images per GPU. Training ran in two stages: 610.27M samples (297,985 steps) at 512-class resolutions, then 49.97M samples (24,400 steps) at 1024-class resolutions, with a global batch of 2,048 throughout.\n\nThe optimizer is split. Muon, with momentum 0.95 and Nesterov, handles the 2D block matrices; AdamW handles everything else; and the Muon updates are RMS-matched to AdamW. The learning rate peaks at 2e-4 with a linear warmup over 1M samples, stays flat to 550M, then decays as 1 minus square root down to 2.5e-5 over the final 60M. The noise schedule runs a curriculum: LogitNormal(0.8, 0.8) for the first 80M samples concentrates training at high noise so structure comes first, then LogitNormal(-0.2, 1.0) widens the distribution so fine detail can be learned. When the model moves to 1024, the noise scale doubles from 2 to 4, and RoPE interpolation aligns the 1024 patch indices back to the 512 coordinate axes.\n\n## The weaknesses, stated plainly\n\nThe model cannot write text or generate posters. That is not an oversight: images of rendered text — infographics, PDFs, book pages, slides — were filtered out of the dataset on purpose, because the lab does not consider that data useful for animation. The prompt upsampler returns a text_requested error for such prompts instead of a caption.\n\nOther limits are listed explicitly: full-body shots, many entities such as six cars or crowds, interactions between several people, and world knowledge. The released failure examples show legs stretched into each other, an extra pair of hands under a person’s legs, and a die where every face shows six dots.\n\nOn prompt-following benchmarks, the model is not the strongest in its own family. On raw prompts, GenEval scores are 0.561 for Linum v2, 0.641 for JiT-DDT and 0.612 for P-JiT. Once prompts are rewritten into the training caption style, the ordering flips: 0.855, 0.808 and 0.817 respectively. DPG-Bench lands at 84.05 for P-JiT on raw prompts and 84.31 on upsampled ones, against 84.79 and 84.54 for Linum v2. The lab’s own explanation is that pixel-space models produce better aesthetic samples and match the underlying image distribution more closely, but appear relatively undertrained on aspects of prompt following such as counting — on the counting task, upsampled GenEval gives 0.80 for Linum v2 against 0.56 for P-JiT.\n\n## Why publishing the bill matters\n\nFirst, the recipe is copyable. It is not just weights and inference code: how the noise schedule is staged, how Muon and AdamW are split, the fact that EMA pauses for 250 steps when the gradient norm exceeds 10 times its rolling median — a spike-aware EMA — these are details that are usually kept private.\n\nSecond, this checkpoint is explicitly a stepping stone. The stated goal is cheaper video generation, because video models almost always start as image models: they have to learn nouns before verbs. Linum notes that a short 720p five-second clip in its v2 model costs 110K tokens of attention, while LLMs train on context sizes of 8K or less for 97 percent of their training regime. Attention is O(n^2) in context length, so cutting token count is where the savings come from. That is why this run was done at all — the lab says the results surprised them for a model this size, and that they are releasing it to encourage others to explore efficient training methods.\n\nThe Hacker News thread drew 95 points and 9 comments, where readers pointed at over-smoothing, an uncanny sheen and mangled limbs, and asked how much cherry-picking went into the showcase. The repository carries 8 stars and a single commit, and both the model card and the README were written by Claude Opus 5.5 on Linum’s behalf, with the model, the training and the code review remaining Linum’s own work.\n\nSo what? The $21,038 is not the headline. What matters is that the bill and the recipe arrived together, which turns \"can I train a usable text-to-image model myself\" from folklore into an engineering question with a documented answer. The more interesting question comes next: if this pixel-space recipe actually transfers to video, is the attention savings enough to drop the price of animation tooling by an order of magnitude?","linum-pjit-pixel-space-text-to-image-21k","2026-11-07T03:08:12Z","2026-11-07T03:08:18.790897Z",254,{"id":92,"title":93,"summary":94,"tags":95,"translations":104,"news_slug":110,"published_at":111,"created_at":112,"image_url":27,"view_count":113},"e7fe0809-52d0-4889-b73d-31bdf58bc8c3","微信WeLM-Image:看图生图改图一体,Turbo版单卡4.98秒出图","腾讯微信团队发布统一图像基座 WeLM-Image:理解、文生图、编辑、多参考合成一体,MoT+MoE,31.7B 存储、约 5.4B 激活,按消费级单卡设计。论文口径三项开源评测第一,Turbo 单张 H20 4.98 秒出 1024 图,权重未放。",[96,99,100,103],{"id":97,"name":98,"slug":98,"description":27,"color":27},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":35,"name":5,"slug":5,"description":27,"color":27},[105],{"id":106,"lang":39,"title":107,"summary":108,"content":109},"886ce088-1a6c-412e-92bf-d7d7a39e16bb","WeChat's WeLM-Image: See, Draw, Edit in One 31.7B Model","WeChat's WeLM-Image unifies seeing, drawing and editing: 31.7B stored, 5.4B active per token; Turbo renders 1024px in 4.98s on one H20.","The default division of labor in image AI has long been: understanding belongs to VLMs, generation to diffusion models, and editing gets its own fine-tuned model. WeLM-Image, which Tencent's WeChat team posted to arXiv on November 5, takes the other road — folding multimodal understanding, text-to-image generation, instruction-guided editing, and multi-reference synthesis into one foundation model, with every design choice bent around a single constraint: it has to run on a single GPU inside WeChat's large-scale consumer applications.\n\n## One sequence, two parameter sets\n\nWeLM-Image processes text, reference images, and target latents as one interleaved sequence. The backbone is a Mixture-of-Transformers (MoT): attention is shared, while understanding and generation each keep independent parameters — a direct attack on the old problem of two tasks interfering with each other. MoE modules on each path add sparse capacity: 31.7B stored parameters in total, with only about 5.4B generation parameters activated per target latent token. Large capacity and low inference cost, demanded at once.\n\n## Data engine and training recipe\n\nTraining relies on a generation-oriented data engine: a knowledge-grounded visual vocabulary paired with large-scale visual clustering to balance semantic coverage and editing tasks, while training-time evaluation steers iterative data refinement. Bilingual annotations and verified editing supervision support progressive training up to native 1536 resolution. Post-training runs in two stages: generation and editing are each optimized with rubric- and checklist-based rewards, then noise-aware on-policy distillation consolidates the two specialized policies back into one model.\n\n## The scorecard, in the paper's own framing\n\nIn the reported overall comparisons, the 1536-resolution model ranks first among the evaluated open-source models on PRISM-Bench, Qwen-Image-Bench, and UnicBench. On speed, the Turbo variant — distilled via MeanFlow few-step training — generates a 1024x1024 image in 4.98 seconds on a single NVIDIA H20 using eight guidance-free steps. The paper lists 16 authors under cs.CV.\n\n## The single-GPU philosophy, and one open question\n\nContrast that with last week's Mistral Large 4 weights landing, where even the 4-bit build still needs about 631.7 GB and a datacenter multi-node setup. WeLM-Image walks the opposite direction: consumer applications at WeChat's scale cannot wait for clusters, so the model yields to a single GPU from day one. This is not the team's first open-research bet — the WeChat vision group previously open-sourced models such as WeVisOne-VL and WeMM. The open question is delivery: as of publishing, the Hugging Face paper page lists no linked model repo, so weights are not out yet — paper first. For anyone building image applications, the thing to watch is not the benchmark scores but whether the unified-base-model-plus-single-GPU route can hold up inside a real app.","welm-image-wechat-unified-model","2026-11-06T13:30:00Z","2026-11-06T13:11:03.041831Z",581,{"id":115,"title":116,"summary":117,"tags":118,"translations":127,"news_slug":133,"published_at":134,"created_at":135,"image_url":27,"view_count":136},"bff1ec03-4e40-4d63-be35-82da33948969","图像生成的长上下文考试:32张参考图拼一张图","华南师大、CMU 等机构联合 KlingAI Research 发布 LongBanana 基准:90 个多参考图像合成任务,每题约 32 张参考图、20 项约束,共 4,670 条专家检查项,已标注被 NeurIPS 2026 接收。RPR 工作流抬高平均分,却在部分框架上拉低通过率。",[119,122,125,126],{"id":120,"name":121,"slug":121,"description":27,"color":27},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",{"id":123,"name":124,"slug":124,"description":27,"color":27},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[128],{"id":129,"lang":39,"title":130,"summary":131,"content":132},"b94d40c4-bb6a-4c71-864e-c85edf039ff0","LongBanana: 32 reference images, 4,670 checks, one image","LongBanana: expert-verified benchmark, 90 tasks, ~32 reference images each, 4,670 atomic checks. RPR agents lift scores but cut pass rates on two harnesses.","Text-to-image models have gotten very good at looking plausible. LongBanana, released by researchers from South China Normal University, Waterloo, the Vector Institute, CMU, Tsinghua, HKUST and HKU together with KlingAI Research, tests a harder skill: given about 32 reference images per task, find the scattered evidence, bind it to the right entities, satisfy 20 constraints, and compose one coherent image. Several authors did the work during internships at KlingAI Research, and the project page marks the benchmark accepted to NeurIPS 2026.\n\n## What the exam tests\n\nLongBanana contains 90 tasks in a 3-by-3 taxonomy — Scene Composition, Visual Appearance and Reference Reasoning — with ten cases per category. Each task carries about 32 reference images, 20 constraints and 50 evaluation items, for 4,670 expert-specified atomic checks in total. Cases are reverse-constructed: experts start from a realistic scenario, build an entity-relation graph, and align reference inventories, synthesis instructions and evaluation criteria to that same graph. At least 30 cases embed two or more hard negatives — visually convincing distractors that are simply wrong to use. The authors call it the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis.\n\n## Scores you can audit\n\nThe same graph grounds evaluation: node, edge and whole-graph obligations become checklist items for Content Fidelity, Relational Binding and Global Constraint Satisfaction. Three critics cover those three categories, with optional OCR, counting, layout and identity tools; every judgment records an evidence trace before aggregation, scored 0, 1 or 2. The human-audited pool spans 90 cases across 10 generation conditions — 900 outputs and 46,700 reviewed labels — and a 15% double-labeled subset shows 89.4% exact agreement with 0.83 quadratic weighted kappa. Image quality becomes an auditable chain of evidence rather than a vibe.\n\n## Agents expose their own weakness\n\nThe companion LongBanana-Gen framework proposes RPR — ReAct plus Plan-Exec plus Reflect — to retrieve evidence, organize generation and repair reference-grounded errors. The results cut both ways: minimal agent harnesses beat direct generation, and RPR lifts checklist averages across the reported harnesses; yet task pass rates drop for OpenCode and Hermes-Agent. Longer reasoning chains create more chances to cite the wrong evidence — more reasoning does not mean the right reference stays bound.\n\n## So what\n\nThe next bottleneck for image generation is not pixels but evidence: retrieval, binding and verification. The dataset is on Hugging Face (Kwai-Kolors\u002FLongBanana), with code and data on the project page. Next time a model \"tops the leaderboard\" on text-to-image, ask one question first: can it pick the right image out of 32 references?","longbanana-multi-reference-image-synthesis-benchmark","2026-11-05T17:05:54Z","2026-11-05T17:05:58.756653Z",240,{"id":138,"title":139,"summary":140,"tags":141,"translations":150,"news_slug":156,"published_at":157,"created_at":158,"image_url":27,"view_count":159},"bf2c4f0f-3911-4227-b741-7252b1875cc7","FrameForge:视频模型跨行图生图,帧距自己挑","FrameForge把统一图生图塞进冻结的视频大模型:时序专家按输入挑虚拟帧距,推理潜变量补上条件到目标的变换,一个LoRA一并通吃主体生成、边缘图生图与指令编辑。论文自评编辑总分 6.12 反超任务专属基线,动作类编辑 6.84 超过表中闭源的 Gemini 2.5,登上 HuggingFace 当日论文榜榜首。",[142,143,146,147],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":144,"name":145,"slug":145,"description":27,"color":27},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":35,"name":5,"slug":5,"description":27,"color":27},{"id":148,"name":149,"slug":149,"description":27,"color":27},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[151],{"id":152,"lang":39,"title":153,"summary":154,"content":155},"99c169bd-364d-4c54-9841-2c90db05e797","FrameForge: Video Models Pick Their Own Frame Gaps","One frozen video model, one LoRA: FrameForge unifies three image tasks. In its table, editing 6.12 beats task-specific; motion 6.84 tops Gemini 2.5.","Unified image generation has a long-standing engineering default: subject-driven generation, canny-to-image, and instruction editing each keep their own adapter. If one model is to cover all of them, a video backbone looks like a natural host — it already learns how frames change into each other. FrameForge ([arXiv:2610.36545](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.36545)), posted on arXiv on Oct 27, is the newest take on this idea: freeze HunyuanVideo1.5, train a single LoRA, and cover all three tasks. Yijing Lin, the lead author, also headed RealGeneral. The authors submitted the paper to the Hugging Face community board only yesterday; it now sits at the top of the daily papers board with 89 upvotes.\n\n## The fixed frame gap is an overlooked assumption\n\nExisting crossovers (RealGeneral, DRA-Ctrl, ChronoEdit) mostly pin the condition frame and the target frame to fixed temporal positions, so every sample shares one \"frame gap\". The paper's observation: this default does not hold. Editing usually changes a small part; subject-driven generation and canny-to-image can overhaul appearance, structure, or the whole scene. The paper's LPIPS distributions show the condition-target variation magnitudes differ clearly across tasks and stay wide within each task — a task label cannot say how much an individual sample should change. Their answer: do not hardcode the gap; let the model pick per input.\n\n## MoTE: three temporal experts, a router picks the gap\n\nThe implementation is a Mixture of Temporal Experts. Three temporal RoPE experts cover three virtual frame gaps with equivalent video-frame spans of 17, 69, and 121. A lightweight router reads the text, condition, and noisy-target representations and picks one expert per sample (top-1); a straight-through estimator keeps discrete routing trainable end to end, and a load-balancing loss prevents collapse. The routing analysis is the most interesting part: over 500 cases and 50 denoising steps, canny — a task with large visual changes — favors the long-interval expert early; subject-driven generation and editing start on the medium interval and shift to the short one as target content emerges, with editing switching earlier and harder. Routing does not tag tasks; it moves with the denoising process.\n\n## CoTR: one unsupervised \"middle frame\"\n\nThe other module handles \"how to change\". CoTR inserts a noise-initialized reasoning latent between the condition and the target, letting it attend to text, condition, and target to gather context step by step. During the intermediate training stage, a stochastic attention mask (probability 0.5) cuts the target's direct attention to text and condition, forcing information through this latent pathway. There is no intermediate-frame supervision anywhere — only the target-frame diffusion loss drives the learning. The ablation also shows more latents are not better: two or three degrade all three tasks, and one is enough.\n\n## The numbers: where it wins and where it loses (all paper-reported)\n\nThe training bill is moderate: backbone fully frozen, a rank-32 LoRA, 182M trainable parameters, 170K samples per task (510K total), 2 A800 GPUs for about 6 days. DreamBench subject-driven overall 0.739, the highest among the compared methods, 1.37% over ChronoEdit (0.729), while subject consistency 0.739 still trails ChronoEdit (0.757). On COCO2017 canny-to-image, SSIM 0.447 is the best in the table, 7.2% over EasyControl; FID 23.47 loses to ControlNet's 20.42 — the authors themselves note FID and SSIM measure different things and should be read together. Editing is the highlight: in the paper's table on GEdit-Bench, the open-source best overall is 6.12, and motion-change editing reaches 6.84, above closed-source Gemini 2.5 (6.72) and Seedream 4.0 (5.53), still below GPT-4o (7.81); overall it also trails the three closed-source systems (7.81-8.01).\n\n## The ablation crack: one fix is not enough, two together overtake\n\nThe most informative row of the ablation table: a shared-LoRA unified baseline drags editing from the task-specific 5.99 down to 4.94 — unification genuinely interferes. Adding MoTE alone recovers to 5.77, CoTR alone to 5.56; neither reaches 5.99. Together they hit 6.12 and overtake the task-specific score. \"Picking the gap\" and \"how to change\" are complementary jobs; fixing only one is not enough. Switching to a HunyuanVideo1.0 backbone keeps the direction (editing 4.38 to 5.00), so the idea does not depend on one backbone. Code and LoRA weights are released, under the Tencent Hunyuan Community License.\n\nWhat is worth remembering is not \"video models won again\", but that \"how much the condition should change\" went from a hand-set prior to an input-dependent, routed quantity. In the paper's comparison table, every open-source method except ChronoEdit — also video-backed — scores between 1.13 and 3.12 on motion editing; what image-native stacks lack may be exactly this sense of change magnitude. When the 6.12-to-8.01 gap closes is a question for the next adapter.","frameforge-adaptive-frame-gap-image-generation","2026-11-02T17:13:17Z","2026-11-02T17:13:25.235269Z",317,{"id":161,"title":162,"summary":163,"tags":164,"translations":172,"news_slug":178,"published_at":179,"created_at":180,"image_url":27,"view_count":181},"2f9ad534-b737-490e-b97e-bf01f2b6d26b","交错生成新基准:闭源旗舰文本赢,图像一致性输给开源","IntReal 用 10 万条真实图文交错样本加七个维度评测统一多模态模型。闭源 NanoBanana2 文本逻辑性 97.49,远超开源 Emu3.5 的 75.17,但图像多样性 83.56、上下文一致性 76.70 反低于开源最优值。SFT 涨文本 29.04% 却让图像美学掉 5.65%,RL 才拉回均值。",[165,166,169,170,171],{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":167,"name":168,"slug":168,"description":27,"color":27},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[173],{"id":174,"lang":39,"title":175,"summary":176,"content":177},"98fcec2e-a43a-44d2-a61f-a312351214ad","IntReal: Closed-Source Wins Text, Loses Interleaving","IntReal benchmarks unified multimodal models on 100K interleaved samples. NanoBanana2 wins text logic 97.49 versus 75.17, but trails on context fidelity.","Scrolling model leaderboards leaves one question unanswered: can the model actually interleave text and images into a coherent piece. A paper posted to arXiv at the end of October ([arXiv:2610.38425](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38425)) breaks that into seven measurable dimensions, and the result is counterintuitive. The closed-source flagship wins big on pure text, yet on the two dimensions that define interleaving, it trails the open-source models.\n\n## Why interleaved generation has been hard to evaluate\n\nInterleaved generation means a model emits a continuous sequence where text and images alternate, like a recipe with a photo next to every step. This is not one image plus one caption: the model has to hold narrative progression, image count, and whether each image matches the preceding context all at once.\n\nThe paper states the situation plainly. Unified Multimodal Models (UMMs) can already jointly generate text and images, but interleaved generation remains \"largely underexplored.\" Two blockers are named: the lack of well-defined interleaved generation tasks with corresponding training data, and the absence of fine-grained evaluation protocols that can reliably diagnose model performance across capability dimensions.\n\nThe gap on the data side is starker. The authors collected a million-scale raw corpus of interleaved text-image content from seven sources that permit public collection and redistribution. Only about 10 percent survived: if any segment of a long interleaved sample is defective, the whole sample goes.\n\n## What IntReal ships: 100K samples plus a 600-sample suite\n\nThe surviving IntReal-100K contains 100,000 samples, 538,452 images, and 146.6M text tokens, averaging 5.38 images and 1466.06 text tokens per sample. The data splits into four categories by how text and images collaborate: Planning (decisions), Knowledge (explanation), Tutorial (ordered steps), and Story (narrative).\n\nThe seven sources contribute 34,771 samples from Xiachufang, 21,138 from Wikipedia, 21,055 from WikiVoyage, 5,055 from Wikitravel, 7,298 from Curious George, 3,949 from Rabbids Invasion, and 6,734 from The Land Before Time. Length distribution tells its own story: Story averages about 8.6 images per sample, the longest interleaved sequences, while Planning averages only 2.42 to 2.68, with text carrying decisions and images visualizing the spatial and temporal changes.\n\nThe evaluation set draws 600 samples by stratified sampling, 150 per category, in two granularities. Holistic evaluation has the model generate the complete interleaved response at once; fine-grained evaluation pins the preceding text-image context and asks only for the next image, so you can localize which step went wrong. The cost is steep: each model has to generate more than 6,000 images interleaved with text.\n\nEach of the seven dimensions gets a dedicated expert model. Text Richness and Text Logicality go to an LLM judge running Gemma4, Image Aesthetic to the RALI IQA model, Image Diversity to DINOv3 similarity between adjacent generated images, Quantity Compliance to plain image counting, Context Fidelity to a VLM judge, and Subject Consistency to DINOv3 comparison against reference images.\n\n## Can the evaluation itself be trusted\n\nThe authors ran three checks. Metrics do not overlap: the absolute pairwise sample-level Spearman correlation between any two dimensions stays below 0.42, with the most strongly correlated pair being Context Fidelity against Subject Consistency. It agrees with people: across five subjective dimensions they collected 750 expert pairwise comparisons, with automatic scores matching expert judgments 83.3 percent of the time on average. It survives a different judge: swapping Gemma4 for the similarly sized open-source Qwen3.6-35B-A3B keeps model rankings correlated above 0.9 (Text Richness 0.905, Text Logicality 0.929, Context Fidelity 0.976).\n\n## The results: closed source wins on text, loses on interleaving\n\nFour UMM families were brought in: Bagel, Emu3.5, SenseNova-U1, and the closed-source API NanoBanana2, which the paper labels Gemini3.1-Flash-Image.\n\nThe outcome is not one model dominating, it is capability profiles that do not line up. NanoBanana2 posts 90.80 on Text Richness and 97.49 on Text Logicality, while the leading open-source Text Logicality score, Emu3.5, reaches 75.17. But on the two dimensions tied to interleaving, it falls behind: Image Diversity 83.56 against an open-source high of 93.31, and Context Fidelity 76.70 against an open-source high of 87.29. The relative deltas in the table put Text Richness 222.33 percent above the open-source high, Image Diversity 10.45 percent below, and Context Fidelity 12.13 percent below.\n\nOne detail in reading that table is worth flagging. The open-source comparison is a per-metric envelope across checkpoints, not one single model, and that envelope already includes checkpoints specifically fine-tuned on this very dataset. The closed-source API is therefore measured against open-source models that had already studied for this test.\n\n## The trade-off post-training exposes\n\nThe authors also checked whether their own metrics can work as training signal. Fine-tuning SenseNova-U1-8B-MoT on IntReal-100K lifts Text Richness by 29.04 percent, Text Logicality by 6.68 percent, and Subject Consistency by 6.53 percent, but costs 5.65 percent on Image Aesthetic and 4.89 percent on Context Fidelity, leaving the seven-dimension mean up just 0.60 percent. That lines up with the widely reported optimization conflict between understanding and generation inside UMMs.\n\nReinforcement learning on top (Flow-GRPO with IQA, DINO, and rubric rewards over 10K carefully annotated samples) pulls Image Aesthetic back from -5.65 percent to -1.34 percent, turns Image Diversity positive at +5.60 percent, and lifts the mean to +2.69 percent, the top score among open-source models.\n\nThe broader conclusion from the authors matters more than the deltas: improving unimodal generation capability generalizes to interleaved generation, so interleaved generation does not require joint text-image training at every stage, and modality-specific optimization is a viable path. Training and evaluation are strictly disjoint: 99.4K training samples, with the 600 evaluation samples removed entirely.\n\n## A few cold-water caveats\n\nThe stated limitations in the paper: high-quality real interleaved data requires aggressive filtering, which keeps the retention rate low, and to keep the experimental setting clean the authors did not explore mixing in unimodal data for post-training, leaving synthetic data augmentation and mixed-data post-training for future work.\n\nSeveral things are worth weighing independently. Human agreement averages 83.3 percent, meaning roughly one in six pairwise comparisons has the automatic score disagreeing with the expert judgment, so this benchmark ultimately has models grading models. The closed-source side is more delicate: the paper explicitly notes that because the system is not publicly disclosed, the authors cannot determine whether the interleaved capability of NanoBanana2 is natively supported or implemented through an agentic pipeline, so reading \"closed source loses to open source\" needs a discount. Licensing is handled source by source: all Story material comes from Curious George, Rabbids Invasion, and The Land Before Time, with each image licensed individually, and user-contributed Xiachufang content requires attribution and is limited to non-commercial use.\n\nOne engineering reality: the whole evaluation runs on a single 80GB GPU such as an A100 80GB or H100 80GB. That is not a high bar, but it explains why the evaluation set stops at 600 samples, since the authors note that scaling further would substantially increase evaluation cost across multiple model families.\n\n## So what\n\nFor anyone building models, the value here is not another leaderboard. It is that \"interleaving ability\" gets decomposed from a vague impression into seven mutually distinct, reproducible dimensions that can even serve as training signal. Being strong at text is not the same as being strong at interleaving, and a lead in the unimodal modality gets erased in text-image alternating scenarios. That is the most counterintuitive finding in these results.\n\nOn October 30, the authors put the dataset, the leaderboard, and the fine-tuned models on Hugging Face simultaneously, with the code repository open alongside. Anyone reproducing this can start running; anyone wanting the training data is welcome to it, since it was prepared for training in the first place.","intreal-interleaved-generation-benchmark","2026-11-01T05:27:08Z","2026-11-01T05:27:30.037114Z",500,{"id":183,"title":184,"summary":185,"tags":186,"translations":193,"news_slug":199,"published_at":200,"created_at":201,"image_url":27,"view_count":90},"b270aaf5-e8a5-40a2-98c7-612acc3573ed","任意顺序图像生成不用重训:适配1.25倍算力换来补图扩图","自回归图像模型因训练时锁死栅格扫描顺序,天生无法补图扩图。这篇 NeurIPS 论文用面向未来的位置编码加顺序课程,把预训练模型改造成任意顺序生成器:适配成本远低于从头重训,并保住 KV 缓存与 vLLM 部署。补图与扩图成绩都超过 MaskGIT,论文也坦承类条件指标有所回落。",[187,188,191,192],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":189,"name":190,"slug":190,"description":27,"color":27},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[194],{"id":195,"lang":39,"title":196,"summary":197,"content":198},"9d45dcaa-171a-4957-995a-430255def714","Unlocking Any-Order Image Generation From Pretrained Models","A pretrained autoregressive image model is turned into an any-order generator, beating MaskGIT at inpainting while keeping KV cache and vLLM serving.","Autoregressive image models carry one hard limitation: they can only paint from the top-left corner to the bottom-right, one token at a time. Because the token ordering is nailed down to a raster scan at training time, the model cannot fill in a patch in the middle, extend the canvas outward, or repaint a single region at inference time. This is not a case of the model failing to learn the capability — the capability was never written into the weights.\n\nThe standard fix is to retrain from scratch, teaching the model a mixture of randomly permuted token orders. The paper cites RandAR, a random-order autoregressive model that needs 300 epochs where the corresponding raster-scan LlamaGen-XL needs 100 — roughly 3x the compute. RandAR's own project page confirms it was trained from randomly permuted token sequences. The other mainstream route, discrete diffusion, also costs about 3x.\n\nA NeurIPS 2026 paper ([arXiv:2610.39293](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39293)) argues that 3x is unnecessary: add two lightweight components to a pretrained model and a finished raster model becomes an any-order generator.\n\n## What actually gets saved is deployment compatibility, not compute\n\nThe paper's comparison table is explicit: RandAR (from scratch) 3x, discrete diffusion 3x, RAR 4x, while this adaptation recipe costs 1.25x on LlamaGen-XL and 1.13–1.33x across four base models.\n\nBut the consequential column is not the training-cost one — it is the three compatibility columns sitting right next to it. RandAR trades 3x for \"cannot serve with vLLM\". Discrete diffusion costs even more and loses both the KV cache and vLLM: its bidirectional attention rescans every unmasked position at each step, which is structurally incompatible with the per-token cache reuse that AR inference frameworks depend on. The adaptation recipe is the only row with all three checked — any-order, KV cache, and vLLM serving — at the price of a 1.25x training budget. The paper notes that for practitioners who already hold pretrained weights, the marginal adaptation cost is over an order of magnitude lower (12x) than training any-order from scratch.\n\nThe technical core is two components, neither of which touches the base model's causal single-stream transformer.\n\nOne component is **future-aware positional embeddings**. Under raster scan, the target position at step t is implicitly known — it is t+1. Once the ordering varies, the same observed context can correspond to entirely different prediction targets: given the rest of one image, predicting a sky token and predicting a face token are two different conditional distributions, so the model receives conflicting gradients simultaneously. The fix is to explicitly tell the model which position to fill next, using target-position-indexed 2D RoPE routing plus an additive embedding table. For LlamaGen-XL that table holds 0.74M parameters, or 0.095%.\n\nThe other is an **ordering curriculum**: rather than switching to random orders in one step, the schedule ramps gradually, laying down a stable optimization path from the pretrained solution to a good any-order basin.\n\n## Three negative results worth more than the headline numbers\n\nThe paper reports FID 6.21 on inpainting against MaskGIT's 10.44, FID 5.84 on outpainting against 12.17, and LPIPS 0.125 on region reconstruction versus MAR-L's 0.174. It is also notably careful with itself: with the Halton schedule, MaskGIT's inpainting FID drops to 7.82, and the paper concedes that part of the original margin came from the competitor's scheduler choice. On outpainting, MAR-L actually has the higher Precision and Recall; with no run-level uncertainty available, the paper calls the two competitive rather than claiming the lead.\n\nThe three counterexamples are what are worth carrying away.\n\n**Remove the future-aware positional embeddings and training loss rises instead of falling.** As the curriculum ramps up the proportion of random orders, the loss curve climbs. That is exactly what the identifiability argument predicts — the gradient conflict is real.\n\n**Remove the ordering curriculum and the model sits half-adapted.** Starting immediately at full random order (α=1) yields FID 5.86 on LlamaGen-L, against 4.24 for the complete recipe. The paper's diagnosis is sharp: Recall barely moves (0.51 versus 0.52), so this is a fidelity collapse rather than a coverage collapse — the pretrained representation gets eroded before order-flexible routing can form.\n\n**The most counterintuitive one: regularizing toward a frozen raster teacher makes things worse.** Adding a KL term (λ=1.0) toward the teacher on raster-ordered examples pushes FID from 4.24 up to 9.62. Preserving the old behavior and learning the new conditional distributions are directly in conflict.\n\nOne engineering detail: the extra embedding table's learning rate has to be scaled 10x relative to the rest of the network to converge stably, otherwise it cannot keep pace with the curriculum.\n\n## The bill still has to be settled\n\nAdaptation is not free. Across four base models, class-conditional FID rises by 1.17–1.93 relative to the corresponding pretrained checkpoint — LlamaGen-L from 3.07 to 4.24, LlamaGen-XL from 2.63 to 3.91, RAR-L from 1.70 to 3.63. The paper volunteers that this cost is \"not necessarily fundamental or monotonic with scale\", and notes that RAR sits closer to the FID floor, where the same nominal degradation is more visible.\n\nOne easily overlooked control: at matched total epochs, adapting a pretrained model clearly beats training from random initialization (LlamaGen-L 4.24 versus 5.60; LlamaGen-XL 3.91 versus 4.81). But the paper reminds us this control matches compute, not schedule — only the pretrained pathway begins with raster-only training.\n\nInference is a separate line on the bill: any-order training makes parallel decoding possible, cutting forward passes by 9x on LlamaGen-XL, at the cost of FID moving from 3.91 to 5.03.\n\n## The boundaries are stated plainly\n\nThe work is confined to class-conditional generation on ImageNet at 256×256. Whether the recipe transfers to text-to-image autoregressive models, or to higher resolutions, remains untested. The editing evaluation covers region completion only, and does not measure instruction following or identity preservation. The adapted model retains vLLM serving through a custom model registration, the code is released under the MIT license, and checkpoints are available under the orderar organization on Hugging Face.\n\nSo what this advances is not a new model but a narrower and harder claim: different factorizations of the same joint distribution can be bridged by one lightweight adaptation, with no need to train a model from scratch. And the three negative results it leaves behind may be worth remembering more than the tidy FID numbers — especially the teacher-regularization experiment that pushes FID from 4.24 to 9.62.\n","any-order-image-generation-ar-adaptation","2026-10-31T07:05:00Z","2026-10-31T07:13:47.710976Z",{"id":203,"title":204,"summary":205,"tags":206,"translations":211,"news_slug":217,"published_at":218,"created_at":219,"image_url":27,"view_count":220},"c03ecfed-2b3f-4831-bbc0-fad67a993917","SpecSlot:像素生成不必让细节都走全局注意力","像素空间生成模型 SpecSlot 把图像词元划分方式改写成预算分配:锚点密度与频谱细分分开算,每个锚点只有最低频那一槽进入全局注意力,其余槽都留在锚点内部做局部交互。论文自测在 ImageNet-512 上 FID 9.14,比同算力对照低 2.84 分且算力更省,文生图 DPG 基准分高出对照 1.06 分。",[207,208,209,210],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},{"id":58,"name":59,"slug":59,"description":27,"color":27},[212],{"id":213,"lang":39,"title":214,"summary":215,"content":216},"eabda3de-e545-484f-9e1a-c9e7cdd12eb6","SpecSlot: Decoupling Pixel Detail from Global Attention","SpecSlot treats tokenization as allocation: only each anchor's lowest-frequency slot joins global attention. The paper reports 9.14 FID on ImageNet.","Pixel-space generative models learn directly from noise to pixels, sidestepping the separately trained autoencoder. The price is that a full-resolution image has to be represented as Transformer tokens. Most such models spend one token per spatial patch, which means a single grid decides three things at once: what the model reads, what each token predicts, and how many tokens join global attention. Want finer detail and the attention bill goes up with it. On a 256×256 image, dropping the patch size from 16×16 to 8×8 takes the token count from 256 to 1,024 and makes the quadratic term of dense attention 16× larger.\n\nThe problem is that coupling. What a token is allowed to communicate globally is decided by where it sits in the grid, not by what it finally produces.\n\n## Rewriting tokenization as budget allocation\n\nA team from Peking University, Nanyang Technological University and SenseTime proposes SpecSlot, and it changes the split itself. The total image token budget is written as N = W × K: W is spatial anchor density, K is spectral subdivision inside each anchor. Every anchor expands into K slots, and each slot owns a fixed, disjoint band of DCT coefficients. A spatial residual is then added on the output side as a local correction. The generative state, the target and the loss all stay in pixel space.\n\nCommunication is assigned by role. In each anchor, the lowest-frequency slot is the carrier, and only that slot enters global attention; every other slot interacts locally inside its own anchor. The schedule is global, global, local, local, repeated three times, giving six global blocks and six local blocks of identical dimensions. The global sequence length therefore tracks W instead of W×K. In the ImageNet-256 configuration with s=16 and K=4, the global sequence holds W+C = 288 tokens, where a scheme that puts every slot into global attention needs 1,056. The corresponding quadratic term drops from 12(WK+C)²d to 6(W+C)²d + 6WK²d.\n\nThe paper summarizes this as tokenization as allocation. It loosens two things that were previously welded together: input density is set by the input grid, and output role is set by the frequency band inside an anchor. Neither has to hold the other hostage.\n\n## The advantage only opens up at 512 resolution\n\nIn the authors' own evaluation, SpecSlot W256\u002FK4 scores 6.59 FID at 129.20 GFLOPs on ImageNet-256, only 0.09 better than the compute-matched JiT-CM control at 6.68. The paper says that gap sits inside the 95% interval across five random seeds, so it is essentially a tie. The real quality leader is still the more expensive JiT-B\u002F8 at 4.34 FID and 218.50 GFLOPs.\n\nAt ImageNet-512 the gap opens. W256\u002FK4 reaches 9.14 FID at 135.64 GFLOPs, 2.84 FID below the compute-matched control at 11.98 while using 3.9% fewer forward FLOPs. Against the costlier JiT-B\u002F16 at 8.45 FID and 219.56 GFLOPs, the gap is only 0.69 FID, at 38.2% less compute and 12.69 against 7.93 img\u002Fs throughput. The authors' explanation is arithmetic: at 512 each anchor covers four times as many pixels as at 256, while the global sequence stays fixed at 288 tokens, so the extra detail stays inside the anchor.\n\nFor text-to-image at 256 resolution, DPG-Bench comes in at 62.21, 1.06 points above JiT-B\u002F8 at 61.15, with forward compute dropping from 218.77 to 129.86 GFLOPs, a 40.6% reduction. The same model scores 26.99 on GenEval, below JiT-B\u002F8 at 29.78. The architecture transfers to text conditioning without structural changes, and the price is giving up ground on a different alignment metric.\n\n## A detail worth not skipping: matching compute is not matching speed\n\nThe routing comparison table says the most about what this work actually is. In the same W256\u002FK4 configuration, GGLL gives 6.59 FID at 129.20 GFLOPs and 9.06 img\u002Fs. The compute-matched dense-global control Dense-global-CM gives 6.81 FID at 128.96 GFLOPs but only 5.80 img\u002Fs. Measured forward GFLOPs are nearly identical, yet wall-clock throughput is far apart. Across sampling budgets of 20, 30 and 50 Heun intervals, GGLL throughput runs 1.47 to 1.49× the control.\n\nOne level further up, Full-global puts all 1,024 slots into global attention and posts the table's lowest FID at 5.84, at a cost of 225.82 GFLOPs, 74.8% above GGLL. So SpecSlot does not take the top score at equal compute. It is cutting 42.8% off the cost of routing everything globally while still holding a better quality than the compute-matched control. It picks the middle operating point, the one constrained by attention, not the leftmost one.\n\n## Mechanistic evidence: the low frequencies really do come back first\n\nThe paper supplies mechanistic evidence. Holding W256\u002FK4\u002FB1536 fixed and changing only which slot gets cross-anchor access: slot 0 gives 6.59, and swapping in slot 1, 2 or 3 gives 6.91, 7.21 and 6.97, all worse. Output-branch probes show the spectral term carries the larger self-energy, while the spatial term's share grows from 0.07 to 0.23 across the flow. Removing either branch at t=0.90 raises velocity MSE by 32.1× and 64.1× respectively, and at t=0.05 by 1.1× and 2.1×. The spectral scaffold is there from the start; the spatial correction is high-frequency top-up that arrives late.\n\nAllocation ablations show that at a fixed slot budget of 2,048, more anchors means a lower score (7.28 for W64\u002FK32, 6.28 for W256\u002FK8, 5.66 for W1024\u002FK2). With anchors fixed at 256, raising K from 2 to 4 to 8 gives shrinking returns, and the last doubling buys only 0.31 FID.\n\n## My read\n\nAll of these numbers currently come from a single origin: the paper, the official repository and the project page belong to the same team, every score is self-reported, and no third party has reproduced it. The paper itself is listed as under review at ICLR 2027. The right way to read it is as a methodological proposal, not as an established result.\n\nThe direction is still worth watching. Most of the past two years of pixel-space generation discussion has circled around how to remove the autoencoder entirely. This work swaps the question for how to decouple detail from communication budget. It does not beat the costlier JiT-B\u002F8 at 256 resolution; what it wins is a non-dominated operating point in the constrained regime. The advantage clearly appears at 512, which hints that the payoff of decoupling grows together with resolution instead of spreading evenly.\n\nFor people building generative systems, the more immediately useful takeaway may be the side observation about controls: two models matched on compute can differ substantially in wall-clock speed. Cutting FLOPs is not the same as cutting latency, especially when what you actually change is the shape of attention rather than only its scale. What to watch next is whether this band division holds at larger scale, and whether the split between low and high frequency still makes sense on temporal data like video.\n\nPaper and code: abstract page [arXiv:2610.40647](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.40647), project page [yangjiongfan.site\u002Fspecslot](https:\u002F\u002Fyangjiongfan.site\u002Fspecslot\u002F), repository [JiongfanYang\u002FSpecSlot](https:\u002F\u002Fgithub.com\u002FJiongfanYang\u002FSpecSlot).\n","specslot-pixel-slot-global-attention-decoupling","2026-10-31T05:20:00Z","2026-10-31T05:16:22.390533Z",226,{"id":222,"title":223,"summary":224,"tags":225,"translations":232,"news_slug":238,"published_at":239,"created_at":240,"image_url":27,"view_count":241},"20f285d4-4a2f-4722-973a-465e3da4c5e1","蚂蚁IDM蒸馏:2步出图,长文本渲染94.25%","蚂蚁在 arXiv 放出 IDM:在 DMD 内部打分之外额外匹配生成器 EMA 特征的均值与标准差,免去外部特征编码器。论文报告在 Z-Image-Turbo 6B 上 2 NFE 时 LongText-CN 达 94.25%、LongText-EN 达 90.20%,消融五项指标全涨。权重尚未发布,数字均为自报。",[226,229,230,231],{"id":227,"name":228,"slug":228,"description":27,"color":27},"471c51be-e620-49df-bd6c-0b5504f53f00","ant-group",{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[233],{"id":234,"lang":39,"title":235,"summary":236,"content":237},"e27a6839-6d8b-4fca-a133-25cc9f840a5e","Ant Group's IDM: Distillation Using the Generator's Own Features","inclusionAI posted IDM on arXiv, matching the generator's own EMA features onto DMD supervision, no external encoder, 94.25% LongText-CN at 2 NFE.","For text-to-image models, the inference bill comes almost entirely from one line item: the number of network evaluations. Diffusion and flow-matching models turn noise into an image through repeated denoising passes, and every extra forward pass adds latency. One-step and few-step generation flattens that curve, at the cost of leaving almost no room for error correction — every character the prompt asks the model to render has to land correctly in the one or two forward passes that remain.\n\nLong-text rendering puts that test on the table. Rows of Chinese and English characters on a poster, a menu, or a book cover have to be both well composed and character-accurate. That is exactly the battlefield Ant Group's inclusionAI picked for its new paper, available at https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39633 .\n\n## The core idea: letting the generator act as its own judge\n\nThe method is called Intrinsic Distribution Matching (IDM), and it builds on the line of distribution matching distillation (DMD). DMD pulls the generated distribution toward a reference distribution using score differences: a pretrained teacher supplies the reference score, minus an auxiliary fake-score network estimating the score of the current generated distribution. IDM follows prior work in estimating that difference with internal velocity branches, which avoids running a separate fake-score network.\n\nWhat is new is an explicit moment constraint. The paper decomposes the reverse-KL field into a Gaussian moment component and a non-Gaussian residual, which identifies a set of statistics that can be constrained directly. IDM retains the full DMD correction while additionally matching feature means and per-coordinate standard deviations — and those features come from the generator's own exponential-moving-average (EMA) copy, requiring no external feature encoder at all.\n\nThat is the most notable part of the whole paper. Earlier approaches bolted on an external feature encoder such as DINOv2 to provide the \"anchor,\" costing extra memory and extra engineering complexity. IDM's answer is that the generator's own EMA copy already carries enough structural information to serve as that anchor.\n\n## The numbers: long-text scores actually pull ahead\n\nThe paper reports that on Z-Image-Turbo 6B at two network evaluations (2 NFE), IDM reaches 94.25% on LongText-CN and 90.20% on LongText-EN, the highest scores among the compared methods at that budget. In the same table, Z-Image-Turbo++ posts 91.62 and 89.88, and APEX posts 90.38 and 88.76. In percentage points, IDM leads Z-Image-Turbo++ by 2.63 and 0.32, and APEX by 3.87 and 1.44.\n\nThat table carries two honest caveats, which the paper states plainly. First, the 8 NFE Z-Image-Turbo reference still scores higher on LongText-EN (91.74), meaning the English side has not been fully caught up. Second, APEX reaches 88.64 on DPG-Bench, above IDM's 86.82. The paper also notes that these published reference models have different training settings and data, so only the ablations under one shared configuration are truly comparable.\n\nThe ablation table is the most persuasive part of the paper. All at Z-Image full-parameter tuning and 2 NFE:\n\n- Without DMD or feature-moment loss: OneIG 52.25 \u002F GenEval 76.24 \u002F LongText-CN 92.70\n- Internal DMD only, no moment loss: 52.87 \u002F 77.46 \u002F 93.87\n- Moment matching with DINOv2 features: 53.00 \u002F 77.75 \u002F 94.13\n- IDM, using generator EMA features: 53.06 \u002F 77.81 \u002F 94.25\n\nAdding the EMA feature-moment loss on top of the internal DMD control yields +0.19 OneIG, +0.35 GenEval, +0.28 DPG-Bench, +0.38 LongText-CN, and +0.41 LongText-EN. What is notable is that the generator's own EMA features do not clearly lose to an external DINOv2 — they edge ahead on most of the metrics. That is the real evidence for the word \"intrinsic\" in the title.\n\nThe speed numbers are equally striking. Full-parameter Z-Image on a single A100 in BF16 takes 0.25 s at 1 NFE and 0.30 s at 2 NFE; Qwen-Image 20B takes 0.35 s and 0.44 s respectively. On throughput, Z-Image reaches 5.18 images\u002Fs at 2 NFE and Qwen-Image reaches 3.28 images\u002Fs. Rank-64 LoRA updates 0.4B parameters and retains most of the full model's GenEval performance. Training cost is disclosed too: 15K steps on Z-Image with 32×H100 80GB, and 10K steps for Qwen-Image LoRA with 8×A100 80GB, at a global batch size of 64.\n\n## A cold shower: the weights are not out yet\n\nThis point has to be made. The badge at the top of the GitHub repository (https:\u002F\u002Fgithub.com\u002FinclusionAI\u002FIDM) says it plainly — \"Model weights: Coming soon\" — and in the Checkpoints section, the full-parameter and r=32\u002Fr=64 LoRA weights for both Z-Image-Turbo 6B and Qwen-Image 20B are all marked \"Coming soon.\" The repository README carries one more self-imposed constraint: \"All values below are reported in the paper. They are not measurements from running this repository.\"\n\nIn other words, right now you cannot run it and cannot reproduce it. The scores in the tables are self-reported, and no independent third-party reproduction exists yet. That does not diminish the methodological value of the paper, but it is a clear wait signal for anyone hoping to use it immediately. The README also warns that base-model weights or APEX weights are not IDM checkpoints and will not reproduce the results above.\n\nTraining-data scale is another variable. The reference data combines ShareGPT-4o, BLIP-3o, and samples synthesized with Qwen-Image-2512 and Z-Image-Turbo, of which 600K come from FLUX-Reason-6M prompts and 5M from text-rendering prompts. The paper contrasts that scale against the 200K poster-prompt samples used by APEX — and a difference in data scale affects final scores on its own, a layer that is easy to skim past when reading comparison tables like this one.\n\n## My take\n\nThe value of this work is not another higher benchmark number. It is a transferable engineering path: swap the bolted-on perceptual anchor for the generator's own EMA representation. One fewer model, one less engineering surface, and long-text scores still went up — a trade that is far more attractive in practice for on-device and self-hosted inference stacks than the percentage points in the paper.\n\nAs for 94.25%, it is better read as a demonstration of \"how far this direction can already be pushed\" rather than \"the current best.\" Once weights ship, third parties reproduce the results, and competitors like APEX rerun under the same data, that is when it becomes meaningful to argue about who is stronger.\n\nThe day those weights land, this recipe is likely the first thing many inference teams copy.\n","idm-intrinsic-distribution-matching-longtext","2026-10-30T11:01:03Z","2026-10-30T11:10:31.569845Z",288,{"id":243,"title":244,"summary":245,"tags":246,"translations":253,"news_slug":259,"published_at":260,"created_at":261,"image_url":27,"view_count":262},"089086a5-bcc1-41a9-b5d2-e8ae17529f25","阿里开源文生图提示词改写器：密度不够反而更糟","阿里把文生图的提示词改写单独立成一个可下载模型，Qwen-Image-PE 的 9B 权重已落在 Hugging Face。论文自测里 9B 综合分 60.28 压过 GPT-5.5 的 60.10，397B 到 62.17。更反直觉的是标度实验：信息密度不够时，改写不是无效而是负收益，低于原始请求。",[247,248,249,252],{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":250,"name":251,"slug":251,"description":27,"color":27},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",{"id":35,"name":5,"slug":5,"description":27,"color":27},[254],{"id":255,"lang":39,"title":256,"summary":257,"content":258},"5a21a4fd-3ec9-42b6-a5da-32622d708b1f","Alibaba Ships a 9B Prompt Rewriter for Text-to-Image","Alibaba's 9B Qwen-Image-PE rewrites text-to-image prompts, beating GPT-5.5 on the team's own benchmark; sparse prompts make it hurt.","A prompt like \"a corgi playing guitar in the rain\" now gets rewritten into a long English description with a suggested 16:9 aspect ratio before the image model ever sees it. Qwen-Image-PE turns that step into a model you can download on its own: the 9B checkpoint sits on Hugging Face, and the inference code lives in the QwenLM\u002FQwen-Image-2.1 repository ([arXiv:2610.33472](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.33472)).\n\n## The target is not prettier, it is design\n\nClassic prompt enhancement just expands a short request into a longer description. Qwen-Image-PE moves the target to four design dimensions, namely lighting, composition, layout and detail, and trains in two stages.\n\nStage one is Caption-Based Reverse Distillation (CRD). Instead of collecting prompts that humans wrote well, the team runs the process in reverse: densely caption images that were designed well, then train the enhancer to predict such a caption from a user request. On 1,000 requests they compare embedding distributions with t-SNE, and the rewritten prompts sit closer to the real training captions. Unigram and bigram Jensen-Shannon divergence drops from 0.229\u002F0.539 to 0.105\u002F0.329, and MAUVE rises from 0.032 to 0.414, with Qwen-3.8-Max as the comparison point.\n\nStage two is Groupwise Design Reward (GDR). A multimodal judge receives the candidate images for one request arranged in a grid, scores all of them jointly under the four design rubrics, and every candidate is scored against the same group of references. The accounting matters: GDR needs D times K judge calls, fewer than pointwise scoring at D times G and far below exhaustive pairwise judging at D times G choose 2.\n\n## The counterintuitive finding: too little information makes rewriting harmful\n\nThree scaling axes stand out. Growing the training set from 12.5K to 100K examples moves Overall from 55.87 to 57.61, a total of 1.74 points, with increments shrinking from 0.81 to 0.63 to 0.30. Swapping the model from Qwen3.5-2B to 397B-A17B moves it from 49.56 to 59.86, a gain of 10.30 points. Moving information density from L1 to L4 moves it from 44.65 to 57.31, a gain of 12.66 points, the densest of the three axes.\n\nThe detail that matters: L1 and L2 land below the unenhanced original request at 54.36, and only L3 and L4 overtake it, by 1.03 and 2.95 points. Below a density threshold the enhancer is not merely useless, it is negative. It takes a request that was already vague and rewrites it into a different kind of vague.\n\n## Three caveats\n\nFirst, every number comes from the team's own evaluation, and Qwen-Image-Bench is Qwen's own benchmark. The 397B variant reaches 62.17 Overall, 7.81 points above the original request, and on the same Qwen-Image-2.1 backbone it leads GPT-5.5 at 60.10, Gemini-3.1-Pro at 60.04 and FLUX 3 Image Klein PE at 59.90. All of these are self-reported paper numbers.\n\nSecond, the 397B model is an internal one, marked with a dagger in the paper's tables. Only the 9B is public. The 9B also clears the baselines above at 60.28 Overall, but its Layout score of 60.72 falls behind GPT-5.5 at 61.60; the 397B lands at 61.11 against 61.15, essentially a tie.\n\nThird, an enhancer improves the input side and still does not outrun the base model. In the same table, GPT Image 2 scores 64.69 Overall as a direct generator, above the 62.17 that Qwen-Image-2.1 reaches with the 397B enhancer attached. The backbones differ, so treat that as a reference rather than a head-to-head. The transfer experiments draw the same boundary. On FLUX 3 Image Klein the native enhancer gets 56.36 while the 9B and 397B reach 57.81 and 59.10. On HunyuanImage-2.1 the native PE-32B manages only 48.59, and Qwen's two variants score 51.04 and 51.66.\n\n## So what\n\nPrompt rewriting is turning into a model category of its own, something you can train, download and deploy separately. Shipping the 9B weights means it no longer requires a closed API, and it also means you own the GPU slot and the inference bill.\n\nThe part worth stealing is the density threshold, not the scores. An enhancer earns its keep only if what it adds is real information. A rewriter that drags the output down when density is low is worse than no rewriter at all.","qwen-image-pe-visual-design-prompt-enhancer","2026-10-28T01:00:00Z","2026-10-28T01:14:43.025888Z",438,{"id":264,"title":265,"summary":266,"tags":267,"translations":272,"news_slug":278,"published_at":279,"created_at":280,"image_url":27,"view_count":281},"2e705005-bc15-4f79-861e-8409799d7ddb","京东LayerEditor:图像编辑直接输出可复用图层","现有图像编辑模型改完只交付一张压平的图,想再动局部只能重生成。LayerEditor 把指令式编辑和 RGBA 图层生成统一进同一 token 序列,配套 10 万配对样本与两个专用基准,在保住编辑保真度的同时产出可拆解复用的图层,登上 HF Daily Papers 榜单前列。",[268,269,270,271],{"id":97,"name":98,"slug":98,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[273],{"id":274,"lang":39,"title":275,"summary":276,"content":277},"8bdb5fb4-2821-45fe-abcf-41698b3d9f92","JD's LayerEditor Turns Image Edits into Reusable Layers","LayerEditor couples instruction editing with reusable RGBA layers in one token model, backed by a 100K paired dataset and two benchmarks.","Anyone who has edited images with AI knows the awkward moment: the model changes exactly one thing, then hands you a flattened bitmap. Want to move the subject or recolor the background? Regenerate and pray. LayerEditor ([arXiv:2610.24738](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.24738)), posted to arXiv on October 17 and submitted under the jingdong organization on Hugging Face, attacks exactly this pain point: while editing the image, also decompose the result into reusable RGBA layers. On October 23 it reached Hugging Face Daily Papers with 116 upvotes, fourth on the day's board.\n\n## One Token Sequence for Canvas and Layers\n\nThe core design is In-Context RGB-Alpha Token Modeling: the edited canvas and an ordered stack of RGB-Alpha layers are jointly represented in a single token sequence. Appearance and opacity are modeled as two distinct, interacting fields; shared attention captures dependencies between the canvas and the layer stack; layer-aware rotary coordinates encode which field a token belongs to and its stacking order, while preserving spatial correspondence. The direct payoff is variable layer counts inside one shared architecture — ask for N layers, get N layers.\n\n## Distillation Keeps the Old Craft Alive\n\nRetargeting a pretrained editing model to emit layers risks degrading its original editing skill during adaptation. Compositional Edit-Prior Distillation (C-EPD) is the cleverest piece of the method: a frozen editing teacher is queried at intermediate states along the student's own denoising trajectory. Training combines direct-output velocity matching with pixel-space supervision on the image recomposed from the predicted layers, and gradients flow back through decoding and alpha compositing. The pretrained editing prior thus transfers to both the edited canvas and the layered output.\n\n## 100K Paired Samples and Two Benchmarks\n\nTraining data comes from LayerEditor-100K: 100,000 paired examples, each with a source image, an editing instruction, an edited target, and an ordered RGBA stack of 2-10 layers. For evaluation the authors introduce Crello-Testset and ScaleEdit-Testset, probing editing fidelity, layer quality, and controllability over layer count and user-specified layer content. The paper reports that LayerEditor maintains strong editing fidelity while producing accurate RGBA layers that support flexible layer-level manipulation and multi-layer decomposition. As of writing, no models or datasets on Hugging Face cite the paper yet — weights and data remain unreleased.\n\n## So What\n\nLayers are the interface between generative models and real production tooling. Designers live in the layer trees of Photoshop and Figma; generative models emit flat pixels. Whoever turns decomposition from a post-processing step into a generation target shortens the path from generated content into real workflows. The C-EPD recipe — freeze a teacher to preserve old capability while learning a new output format — generalizes beyond image editing. Of course, a paper is not a tool: without released weights, this is still the eve of the demo. For anyone building design tools or AIGC pipelines, this line of work belongs on the watchlist.","layereditor-rgba-image-editing-layers","2026-10-24T13:10:00Z","2026-10-24T13:07:52.021851Z",402,{"id":283,"title":284,"summary":285,"tags":286,"translations":291,"news_slug":297,"published_at":298,"created_at":299,"image_url":27,"view_count":300},"bbbda7ff-b1c7-4142-b37f-2b3fd0547e02","Fibo 2.0 开源:生成编辑合一,登顶开源编辑榜","Bria 开源图像模型 Fibo 2.0-dev,首次把图像编辑与文生图统一进单一模型;官方引 Artificial Analysis 10 月 20 日榜单,称其在开源权重模型中图像编辑第一、文生图第二。基于结构化 JSON 描述与全授权数据训练,主打可控编辑,另有蒸馏版 Edit Turbo 同期发布。",[287,288,289,290],{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[292],{"id":293,"lang":39,"title":294,"summary":295,"content":296},"f764c82e-1203-4109-8949-738b7f23f948","Bria Fibo 2.0-dev: One Open Model for Both Editing and Generation","Bria's Fibo 2.0-dev unifies image editing and text-to-image in one open model, ranked #1 in open-weight editing on Artificial Analysis per its card.","This year's image-model arms race has mostly been about fidelity: whose photos look truer, whose leaderboard score is higher. But teams that actually put these models into production pipelines will tell you fidelity is just the entry ticket — the hard part is control. Change one prompt clause and the face shifts, the composition drifts, the lighting re-arranges itself. Bria has always taken the other road: compete on control, not imagination. On October 18 the company put its next-generation model, Fibo 2.0-dev, on Hugging Face — for the first time folding image editing and text-to-image generation into the same model.\n\n## One model, two jobs\n\nThe headline feature of Fibo 2.0 is unification. A single model handles text-to-image generation as well as instruction-based image editing, and it supports masked edits, multi-reference inputs and iterative refinement. The model card stresses this is a first for the FIBO family — editing and generation are no longer separate stacks, one deployment covers both. The release ships two variants. Fibo 2.0-dev is the full, non-distilled base model that runs classifier-free guidance for maximum quality, positioned as the starting point for professional fine-tuning and research. Fibo 2.0 Edit Turbo is a distilled few-step, low-latency editing variant with the same architecture, aimed at production. The official roadmap also lists a multi-turn editing LoRA: chained localized edits where each round touches only what it should, while the rest of the image stays pixel-stable.\n\n## Structured descriptions are the moat\n\nThe core idea behind the FIBO line is replacing natural language with structured description. The model is built on what Bria calls the VGL (Visual GenAI Language) paradigm: lighting, composition, color, camera parameters, depth of field and every object in the scene become separate, addressable fields in JSON. Generation becomes \"render this blueprint\"; editing becomes \"apply a diff to the blueprint.\" Training data comes exclusively from Bria's roughly one-billion-image licensed dataset, paired with long structured captions of up to a thousand-plus words. The direct payoff is what the card calls native disentanglement: change only the camera angle, or swap day for night, and the rest of the scene does not drift. For production workflows that need repeatable, consistent output, this is the real requirement.\n\n## First on the leaderboard, but two buckets of cold water\n\nThe official model card cites Artificial Analysis rankings as of October 20: among open-weight models, Fibo 2.0 places first in image editing and second in text-to-image. Note this is the vendor quoting the benchmark, not an independently re-verified result. The card also claims clear gains in in-image text rendering, material texture and human anatomy (hands, faces, poses) — likewise official self-reporting until third-party evaluations land.\n\nTwo caveats. First, the license: the weights ship under CC BY-NC 4.0, non-commercial use only, and you must log in and accept terms before downloading — strictly speaking this is \"open weights,\" not grab-and-go. Commercial use requires Bria's own licensing terms. Second, the controllable route has a learning curve: you have to learn to write structured descriptions first, and whether the reproducibility you gain is worth the workflow overhaul depends on whether your business genuinely needs field-level control.\n\n## So what\n\nThe competitive axis of open image models is shifting. With fidelity now near-indistinguishable to the eye, controllability, licensing compliance and editing capability are becoming the reasons people pay. Bria's \"licensed data only plus structured captions\" combo digs a genuinely differentiated moat, and the pull is real for copyright-sensitive brand and e-commerce customers — yet CC BY-NC 4.0 means most teams can evaluate but not ship. The thing to watch is when the multi-turn editing LoRA lands: if conversational chained editing becomes standard, the boundary between \"image model\" and \"image editor\" may genuinely disappear.\n\nSource: [briaai\u002FFibo-2.0-dev model card](https:\u002F\u002Fhuggingface.co\u002Fbriaai\u002FFibo-2.0-dev)","bria-fibo-2-0-unified-editing-generation","2026-10-21T21:08:11Z","2026-10-21T21:08:16.569651Z",632,{"id":302,"title":303,"summary":304,"tags":305,"translations":310,"news_slug":316,"published_at":317,"created_at":318,"image_url":27,"view_count":319},"63299a5b-f3b7-40c4-9985-0132e0cbffac","多轮图像编辑漂移的解法：MT Forcing 用自回归扩散把单轮质量带进多轮","商汤等团队提出 MT Forcing 训练范式，用单轮分布匹配蒸馏加教师强制引导，压制自回归扩散在多轮图像编辑中的误差累积。在 MTEdit-Bench 上，八轮编辑的指令遵循从 8.47 提到 9.66，内容一致性从 7.09 提到 8.40，像素漂移从 6.34 降到 0.74。模型权重已开源。",[306,307,308,309],{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[311],{"id":312,"lang":39,"title":313,"summary":314,"content":315},"b31ff89a-226d-48de-8031-4964e4ea1859","Stabilizing Multi-Turn Image Editing with MT Forcing","SenseTime-led MT Forcing curbs drift in autoregressive multi-turn image editing: eight-turn instruction following 8.47 to 9.66 on MTEdit-Bench.","Rarely does a finished image stop at a single edit. You add an object, adjust the lighting, then refine a detail — and the real difficulty is keeping the parts you were already happy with from being quietly destroyed by the next pass. Autoregressive diffusion models drift in exactly this way during multi-turn editing: textures blur little by little, the subject shifts slightly, and after a few turns the image is unrecognizable.\n\nA recently released paper proposes a training paradigm aimed squarely at that failure mode. MT Forcing, from SenseTime Research with HKUST MMLab, Stanford and Tsinghua AIR, trains on the model's own multi-turn rollouts while drawing supervision from a single-turn teacher line — the teacher always starts from the original image, applies every accumulated instruction in one pass, and its output distribution corrects the student. The paper and project page are available at [arXiv:2610.25404](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.25404) and on the [project site](https:\u002F\u002Fhouyuanchen111.github.io\u002Fmt-forcing.github.io\u002F).\n\n## How the drift happens\n\nThe paper splits multi-turn editing failures into two categories: one is unstable visual quality, where artifacts accumulate across turns and details degrade; the other is inconsistent content, where regions that should be preserved and earlier edits get changed unintentionally. When a model continuously edits its own outputs, small errors compound step by step.\n\nThe framework builds on unified multimodal models (UMMs) with autoregressive diffusion. The authors describe two complementary training signals, one for each failure category.\n\n## Two signals: one holds quality, one holds consistency\n\n**Single-Turn Distribution Matching Distillation (Single-Turn DMD)** targets visual quality. The student edits its own generated history one turn at a time, while the teacher starts from the original image, composes all instructions accumulated so far in a single pass, and provides distribution-level guidance. That way the teacher carries no error from one turn into the next, avoiding recursive error accumulation. The project page summarizes it as a single-turn teacher present at every turn.\n\n**Teacher-Forcing Guidance** targets cross-turn content consistency. It supervises history-dependent transitions along real multi-turn trajectories: reference trajectories guide the warmup phase, and the signal continues as a regularizer during distillation. The authors state explicitly that per-turn distribution matching alone cannot guarantee cross-turn consistency — which is exactly why the second signal is needed.\n\nBoth signals act on the same student model, each covering one side of the problem.\n\n## The numbers across eight turns\n\nThe disclosed figures come from 100 eight-turn sessions in MTEdit-Bench, scored as eight-turn averages with SenseNova-U1.5 as the base model:\n\n- **Instruction following (IF)**: 8.47 base → **9.66** with MT Forcing (out of 10)\n- **Content consistency (CC)**: 7.09 → **8.40**\n- **Drift resistance (DRS)**: 3.83 → **4.30** (out of 5)\n- **Matched-point displacement**: 6.34 px → **0.74 px** (lower is better)\n\nThe project page also spells out the evaluation protocol: eight-turn means weight every turn equally, and IF and CC follow the benchmark's failure-triggered scoring rule, where turns after a failure score zero. Pixel metrics measure mean displacement between matched points, with lower values indicating better spatial alignment, and each metric is interpreted separately.\n\nLooking at turn 8 alone makes the gap starker: 7.68 against the base model's 3.22, a 4.46-point margin. The project page also provides a fifteen-turn editing comparison demo.\n\n## Reading it\n\nThe interesting part here is not any single score but a training-signal design problem that tends to get overlooked: in multi-turn settings, how the teacher is defined determines whether error compounds. Most methods assume teacher and student sit at the same turn, so the teacher climbs forward carrying its own accumulated error. MT Forcing instead sends the teacher back to the original image and folds all instructions into one pass, which in effect makes the supervision signal immune to historical error.\n\nTwo caveats are worth stating plainly. First, the repository has not released code yet — it is officially marked as coming soon, with training code, inference code and pretrained checkpoints all absent, so anyone wanting to reproduce immediately will have to wait. Second, the reported numbers are team-published results on MTEdit-Bench, the authors' own benchmark; third-party replication has not appeared so far, which is the discount one should apply when reading evaluation work of this kind.\n\nMulti-turn editing is moving from demonstrable to deliverable — users will not accept an image that no longer looks like itself after five edits. Whoever trains \"remembering what came before\" into the model holds the actual entry point for interactive visual creation.","mt-forcing-multi-turn-image-editing","2026-10-20T05:06:17Z","2026-10-20T05:06:21.878777Z",490,{"id":321,"title":322,"summary":323,"tags":324,"translations":331,"news_slug":337,"published_at":338,"created_at":339,"image_url":27,"view_count":340},"7010d17a-2089-48fe-88b5-cd1eaef0de72","WorldCast基准:GPT-Image-2.5拿72.8分,最强开源只有20.1","美团实习期间提出 WorldCast 基准:320 案例专考图像视频模型多步世界状态预测,12 模型评测,GPT-Image-2.5 得 72.8,Nano Banana Pro 42.8,最强开源 SenseNova-U1 仅 20.1;Agent 循环把最高分推到 82.4,瓶颈正转向语言模型的规划推理。",[325,326,327,330],{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":328,"name":329,"slug":329,"description":27,"color":27},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",{"id":35,"name":5,"slug":5,"description":27,"color":27},[332],{"id":333,"lang":39,"title":334,"summary":335,"content":336},"47b5fdde-18e6-4856-bb0e-6a57e5d925a8","WorldCast: GPT-Image-2.5 Scores 72.8, Best Open Model 20.1","WorldCast tests multi-step world state prediction on 320 cases: GPT-Image-2.5 scores 72.8, Nano Banana Pro 42.8, and the best open model only 20.1.","Asking an image model to \"draw the future\" is far harder than asking it to draw well. Given a few observation images and a manual, the model must render each successive world state in order — assembly steps, physical changes, and rule inference all have to be right. This task, multi-step world state prediction, demands tightly entangled understanding and generation, plus the world knowledge to translate human action sequences, physical transformations, and abstract visual reasoning into coherent frames. A study completed during an internship at Meituan proposes the WorldCast benchmark, organizes model capabilities into a four-level hierarchy (L1–L4), and argues that higher-level tasks remain underexplored despite rapid model progress. WorldCast targets exactly that L3–L4 band. [arXiv:2610.20731]\n\n## 320 Cases, Three Scoring Dimensions\n\nWorldCast contains 320 manually designed cases across eight task categories, requesting 1,352 visual states in total. Categories are not split by \"what to draw\" but by where the information governing a future process is encoded and what reasoning is needed to turn it into state transitions: physical simulation, adaptive task completion, visual goal realization, in-context procedural learning, and more. One L4 case asks the model to read a furniture assembly manual, plan the order itself, and render every intermediate state while preserving the work surface and component appearance. On the project page's shown result, GPT-Image-2.5 completes the key milestones while Wan 3.0 fails to follow the assembly order.\n\nScoring uses three complementary dimensions: Task Completion (TC, whether the sequence visibly realizes the task), Causal Plausibility (CP, whether transitions follow physical laws and logical rules), and Visual Consistency (VC, whether identity, count, geometry, and material persist). GPT-5.5 acts as the judge on a 0–3 scale; the final score is a weighted, rescaled combination.\n\n## The Leaderboard: A Bigger Cliff Than Expected\n\nTwelve image and video generation systems were evaluated under the same protocol. GPT-Image-2.5 System (paired with GPT-6 Astra) scores 72.8, which the paper describes as early-stage visual state prediction capability on real-world tasks. Nano Banana Pro reaches only 42.8. The best open-source model, SenseNova-U1, scores 20.1 — a gap of more than 3.5x between the top closed system and the best open weights.\n\nThe divergence grows with difficulty: GPT-Image-2.5 scores above 60 on all five L4 categories and retains 71.6 on L4, only 2.2 below its L3 score. SenseNova-U1 drops from 24.4 at L3 to 15.0 at L4. The higher the level, the wider the gap. Failure modes are concrete: unintended rotation, wrong reflections, erroneous density, shape drift, and incorrect assembly order.\n\n## Three Findings\n\n**Strong generators draw on their language backbones.** Pairing Nano Banana Pro with GPT-6 Astra as the enhancer raises its overall score from 42.8 to 61.2, yet it still trails the GPT-Image family. Conversely, swapping GPT-Image-2's backbone from GPT-5.5 to GPT-6 yields only marginal gains, with a decline in Physical Simulation. The ceiling of generative capability is welded shut by the planning and reasoning capacity of the language model.\n\n**Output format is a hidden variable.** Switching GPT-Image-2.5 from interleaved to grid output cuts its overall score by 22.7, with Logic and Rules falling from 74.3 to 23.0. Across four generation steps, weaker models degrade as context accumulates, while GPT-Image-2.5 stays stable.\n\n**Agentic harnesses raise the ceiling further.** The researchers built two Codex-based harnesses with the same generation budget: an inspect-and-revise loop without coding tools lifts GPT-Image-2.5 from 72.8 to 80.2 and GPT-6-enhanced Nano Banana Pro from 61.2 to 72.5. A coding-first loop that constructs and inspects intermediate visual artifacts before generation pushes the best overall score to 82.4, with the largest gains on Visual Goal Realization, Logic and Rules, and Physical Simulation — code makes intermediate states explicit.\n\n## So What\n\nThe real signal of this benchmark is not \"GPT-Image-2.5 won again,\" but that the image-generation race is shifting from \"drawing it well\" to \"reasoning it right\" — and the latter is the language model's home turf. The open-source side's 20.1 overall and 15.0 at L4 suggest open weights have not truly entered the high-level world-modeling game; catching up may require not a bigger diffusion backbone but better planning and reasoning layers. The dataset is on Hugging Face (wenhaoyan77\u002FWorldCast), and the project page (teal024.github.io\u002FWorldCast) carries the full leaderboard and case gallery.\n\nWhen \"looks good\" is saturated, the next dividing line is: does your model know what the world should look like next?","worldcast-benchmark-world-state-prediction","2026-10-19T15:09:58Z","2026-10-19T15:10:02.044926Z",379,{"id":342,"title":343,"summary":344,"tags":345,"translations":350,"news_slug":356,"published_at":357,"created_at":358,"image_url":27,"view_count":359},"3bf416bb-a989-4b02-8341-dbafc7c96c48","MOSS-Image 开源:5B 双语生图,文字排版可编程","OpenMOSS 放出 5B 生图基座 MOSS-Image-Base 与 SFT 版,Apache 2.0。模型原生支持中英双语提示词与 4K token 长输入,主打图内文字渲染,可用 JSON layout 精确控制文字与物体位置;官方未公布 benchmark,衍生微调已现。",[346,347,348,349],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[351],{"id":352,"lang":39,"title":353,"summary":354,"content":355},"3487d92d-fbd2-4a04-a74b-2b8d58e2928d","MOSS-Image Open-Sourced: 5B Bilingual Base with JSON Layout","OpenMOSS ships a 5B Apache-2.0 DiT image base: bilingual prompts, 4K-token inputs, in-image text rendering, JSON layout control; no benchmarks yet.","On October 16, the OpenMOSS team announced Base and SFT versions of MOSS-Image on GitHub, with weights simultaneously published on Hugging Face, both weights and inference code under Apache 2.0. In an open image-generation field crowded with new FLUX and Qwen-Image releases, this 5B-parameter DiT base takes a differentiated route: instead of chasing general aesthetics, it focuses on in-image text rendering and layout control, with native bilingual prompt support.\n\n## A 5B bilingual base\n\nMOSS-Image-Base is the base model of the MOSS-Image family, positioned by the team as a foundation for developers and researchers to fine-tune downstream rather than a finished consumer product. Core specs: a 5B-parameter DiT backbone, bfloat16 for the transformer and text encoder, float32 VAE, and a bundled VAE that directly reuses the FLUX.2 small decoder — the README acknowledgements name the Diffusers, Qwen, and FLUX teams, a fairly honest disclosure of its technical lineage.\n\nThe prompt side has two practical designs. First, native English and Chinese support: the official quick-start example is a Chinese handwritten note — a fridge sticky note reading \"remember to buy milk and eggs\". Second, the team says prompts up to 4K tokens are accepted, and tag sequences can be fed directly.\n\n## JSON layout: typesetting as a programmable input\n\nThe genuinely distinctive part is layout control. MOSS-Image accepts a JSON structure as the prompt: `summary.caption` describes the overall scene, the `items` array declares elements one by one, each with `label`, `caption`, and `bbox`, and any text to render goes in the `text` field. Coordinates use `yxyx` format with absolute pixel coordinates centered on the canvas — on a 1024×1024 canvas each axis runs from -512 to +512. A request like \"this line of text on a sticky note, the note on the fridge's upper-left corner\" thus turns from natural-language guesswork into precisely editable JSON: change the text, the description, or the box position, and the composition adjusts in a structured way; controlled editing on the same layout is also supported.\n\nThe official sampling configuration is fully disclosed: flow-matching Euler, 50 steps, CFG 3, shift 5, the negative branch skips layer 22 (zero-based) by default, and output dimensions must be multiples of 16. Dependencies are just diffusers 0.38+ and transformers 5.3+; a local custom pipeline runs in a few lines.\n\n## Fast ecosystem growth, absent benchmarks\n\nSince launch, the Hugging Face page already lists 16 fine-tuned derivatives, 9 adapters, 2 quantized versions, and 12 Spaces demos, while the GitHub repo sits at 485 stars. A native ComfyUI plugin is available in ComfyUI Manager, and integration into the official ComfyUI repository is in progress. The repo also ships a \"prompt enhancer\" agent skill, designed for GPT-6-family agents, that compiles vague image ideas into detailed captions or layout JSON. Another detail worth noting: the official highlights explicitly list heterogeneous GPU training — the model was trained on alternative GPU platforms, a non-trivial signal under the current compute landscape.\n\nThe cold water is equally clear. OpenMOSS did not publish benchmark results or a supported-resolution list alongside the release, so horizontal comparison beyond parameter count is not yet possible; \"strong text rendering\" remains a self-reported claim until third-party reproduction data appears.\n\n## So what\n\nThe value here is not \"yet another open 5B image model\", but that MOSS-Image turns in-image typesetting into structured input — posters, e-commerce detail pages, and captioned memes, the everyday needs of Chinese-language scenarios, are exactly the gap that quite a few Western open models never seriously covered. Base plus SFT dual releases, Apache 2.0, and low fine-tuning barriers form a standard \"let the community grow the ecosystem\" playbook. The 5B model's true standing will only be clear once third-party benchmarks and LoRA tests arrive; but the bilingual text-rendering wedge is already enough for it to hold uncrowded ground.","moss-image-bilingual-text-rendering-base","2026-10-18T19:08:44Z","2026-10-18T19:08:57.779589Z",1112,{"id":361,"title":362,"summary":363,"tags":364,"translations":369,"news_slug":375,"published_at":376,"created_at":377,"image_url":27,"view_count":378},"b88d9e4c-1450-4349-a3cd-c2f8c0ce50eb","FLUX 3 Klein 开源:5.6B,8GB 显存能跑","Black Forest Labs 开放 5.6B 生图模型 FLUX 3 Image Klein 权重,支持文生图、编辑与最多 10 张参考图组合,继承 4K 生成与边界框控制。INT8 版最低 8GB 显存可跑,蒸馏版 RTX 5090 约 4 秒出 1MP 图,官方内部评测开放权重排名第一。",[365,366,367,368],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[370],{"id":371,"lang":39,"title":372,"summary":373,"content":374},"3e517942-35f5-4704-9f40-83f3e920352e","FLUX 3 Klein Open Weights: 5.6B Image Model Runs on 8GB VRAM","BFL open-sources FLUX 3 Image Klein, a 5.6B text-to-image and editing model running on 8GB VRAM, first in BFL's internal open-weights rankings.","On October 15, Black Forest Labs (BFL) shipped the second open-weights release from its FLUX 3 family: FLUX 3 Image Klein. The first was FLUX 3 Action, a 7B model for robot control; this one is for images. 5.6B parameters, consumer-GPU friendly, generation and editing in a single model — BFL keeps pushing the \"flagship quality, local deployment\" line further down the hardware stack.\n\n## Small transformer, clear division of labor\n\nThe 5.6B count refers to the image transformer itself; the text encoder is separate — Alibaba's Qwen3-VL-4B-Instruct, per the Hugging Face model card. The model takes a prompt plus up to 10 reference images and returns one image: text-to-image, single-reference edits and multi-reference compositions all run through the same weights.\n\nTwo checkpoints shipped: a base model for fine-tuning and LoRA training, and a step-distilled model for low latency. Quantization comes in FP8, INT8 and NVFP4 builds developed with NVIDIA. Generator weights are 11.2 GB in BF16 and 5.6 GB in FP8; with the INT8 build the model runs on RTX GPUs with as little as 8GB of VRAM via ComfyUI or Diffusers. On speed, the distilled model with INT8 generates a 1MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070; on GB300, NVFP4 comes in under 0.6 seconds.\n\n## Vendor evals: the small one ranks first\n\nIn BFL's internal evaluations, Klein ranks first among open-weights models in text-to-image, multi-reference generation and image editing, and rates above the closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro]. It is the smallest model in that comparison: 5.6B against 7B for Qwen-Image 2.1 and 12.9B for Krea 2. Versus the previous FLUX.2 [klein] 9B, it gains 91 Elo in text-to-image and 70 in editing, with 38% fewer parameters. Caveat: these rankings all come from the vendor's own evals, with no independent third-party replication yet — but the \"fewer parameters, higher score\" direction is consistent with FLUX 3's pretraining recipe, which shares signals across images (spatial detail), video (motion and physical dynamics) and audio (events and interactions).\n\nKlein also inherits FLUX 3 Image's core capabilities: native 4K generation, pixel-perfect editing, and bounding boxes that control where elements land or confine edits to a defined area.\n\n## Ecosystem and license: the gate is commercial\n\nDay-one support is broad: ComfyUI, Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni; LoRA training via Ostris's AI Toolkit and recipes in NVIDIA NeMo AutoModel. The license is the FLUX Kommunity License: free commercial use for individuals and companies under 5 million US dollars in annual revenue, plus research and non-commercial use; larger organizations need self-hosting licenses or the API. API pricing for the distilled model starts at 0.009 US dollars per 1K images. On safety, the checkpoints were red-teamed internally and evaluated by external partner Cinder; BFL says they show substantially fewer vulnerabilities than other popular open-weights models.\n\n## So what\n\nThe story here is not leaderboard points but the dropping barrier to local image editing: 8GB of VRAM covers a large slice of laptops on sale today, and a 4-second distilled generation enters interactive territory. Add Qwen3-VL as the text encoder and Chinese-prompt friendliness is worth watching. The fine print that matters is the revenue line — a free lunch for individuals and small teams, a negotiation for enterprise self-hosting. The next question: will native 4K plus 10 reference images push local workflows from generation all the way to editing?\n\nSources: the [BFL release post](https:\u002F\u002Fbfl.ai\u002Fblog\u002Fflux-3-image-klein) and [Hugging Face model card](https:\u002F\u002Fhuggingface.co\u002Fblack-forest-labs\u002Fflux-3-image-klein); independent coverage at [VentureBeat](https:\u002F\u002Fventurebeat.com\u002Ftechnology\u002Fblack-forest-labs-opens-flux-3-image-klein-weights-with-free-commercial-use-limited-to-smaller-businesses).","flux-3-image-klein-open-weights","2026-10-17T13:20:00Z","2026-10-17T13:10:47.851121Z",1223,{"id":380,"title":381,"summary":382,"tags":383,"translations":387,"news_slug":393,"published_at":394,"created_at":395,"image_url":27,"view_count":396},"162c36d7-37de-4906-b0dd-cd5ec6f9f455","微软擦除基础模型LensErase:一图一掩码,万物皆可擦","微软联合上海交大等高校发布掩码式擦除基础模型LensErase:一图一掩码,无需文字提示即可抹掉对象、文字、水印及其阴影反射。模型基于文生图底座Lens,通过Mask-Aware RoPE注入掩码信息,在1.01亿组配对数据上预训练,再经SFT与契约引导强化学习精调;六个公开基准与自建500案例评测中均取得大幅提升。",[384,385,386],{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[388],{"id":389,"lang":39,"title":390,"summary":391,"content":392},"d7436c05-cabf-4733-8112-5fb2b9a0a1ad","LensErase: Microsoft's Erase-Anything Foundation Model","Microsoft-led LensErase removes any object, text or watermark from one image plus one mask, pretrained on 101M pairs for large gains on benchmarks.","The most thankless step in image editing is erasure. Generative models keep getting better at adding things to a picture, but removing an object together with its shadow and reflection, filling the revealed background seamlessly, and leaving the surrounding content untouched pixel-for-pixel has long lacked a unified solution. On October 13, a Microsoft-led team with researchers from Shanghai Jiao Tong University, Nankai University, USTC, the University of Sydney and Peking University released LensErase on arXiv, framing erasure as a foundation-model task ([arXiv:2610.18072](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.18072)).\n\n## One image, one mask, no prompt\n\nLensErase compresses the interaction to the minimum: an input image and a selection mask, with no text description required. Covered scenarios include general objects, visual text, watermarks and layered posters, multiple targets and semi-transparent content in a single pass, plus associated visual effects such as shadows and reflections — exactly where traditional retouching tools tend to fail.\n\n## 101 million paired examples, then reinforcement learning\n\nTo make a model that can erase anything, the team first built the data: LensErase-100M contains 101.07 million source-mask-target triplets spanning objects, watermarks, large masks, visual text and posters. The base is their own text-to-image model Lens, whose generative prior is adapted to mask-based erasing. To feed mask information into the network, the authors propose Mask-Aware RoPE, a 3D rotary positional encoding that injects the mask directly into source-image tokens without a separate mask token sequence. The model then goes through supervised fine-tuning and reinforcement learning with contract-guided, region-aware rewards; at inference, TraceFusion uses image residuals to identify the effective edit region and keeps everything outside it untouched.\n\n## Six public benchmarks plus a 500-case internal bench\n\nEvaluation is doubled up: six public benchmarks plus the proposed LensErase-Bench (500 cases, no paired ground-truth required), scored both by reference-based objective metrics and reference-free VLM evaluation. The bench breaks out Removal, Inpainting, Preservation and Cleanliness, plus an Overall Score and HiFi Removal. The paper reports substantial improvements across both evaluation families, and the project page adds qualitative comparisons against GPT-Image 2.5 and Flux2-Erase. The code repository is up ([GitHub](https:\u002F\u002Fgithub.com\u002FLensFamily\u002FLensErase)) under an MIT license, currently marked \"Code coming soon\" — no model weights are hosted yet, and the count of Hugging Face models citing the paper is still zero.\n\n## So what\n\nOnce generation has been standardized by foundation models, erasure is becoming the next standardized editing primitive. Notably, LensErase lists watermarks as a first-class supported target: while OpenAI and Anthropic race to embed provenance watermarks into AI-generated content, an erase-anything model quietly makes watermark removal a baseline capability. Which way this cat-and-mouse game tips is worth watching closely.","lenserase-erase-anything-foundation-model","2026-10-15T21:08:38Z","2026-10-15T21:08:54.529996Z",466,{"id":398,"title":399,"summary":400,"tags":401,"translations":406,"news_slug":412,"published_at":413,"created_at":414,"image_url":27,"view_count":415},"59d14de3-0e17-4202-8e7d-ad0bc51e3471","Qwen-Image-2.1-Turbo开源:8步去噪出图","阿里千问开源 Qwen-Image-2.1-Turbo:同为 7B 架构,40 步去噪压到 8 步,文生图与修图都支持,2K 输出、透明背景全保留。同日 API 上线,Turbo 每张 0.1 元。权重走 Qwen Research License,商用需授权;社区实测皮肤纹理过度锐化,画质待独立评测。",[402,403,404,405],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":250,"name":251,"slug":251,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[407],{"id":408,"lang":39,"title":409,"summary":410,"content":411},"4f5d9567-338f-4642-8f93-93818fdda018","Qwen-Image-2.1-Turbo Open-Sourced: 8-Step Image Generation","Qwen ships Qwen-Image-2.1-Turbo: same 7B architecture, denoising cut from 40 steps to 8. APIs cost 0.1 RMB per image; weights stay research-licensed.","On October 9, the Qwen team updated the Qwen-Image-2.1 GitHub repository with two announcements: the Turbo checkpoint is available for download, and the Pro and Turbo APIs are now live on Alibaba Cloud Model Studio. That is less than three weeks after the base 2.1 open-weight release on September 20 — a notably fast cadence.\n\n## From 40 Steps to 8\n\nTurbo keeps the same 7B visual generation architecture as Qwen-Image-2.1; the change concentrates on sampling efficiency. The original model's official examples run 40 denoising steps, while Turbo cuts that to 8 — an 80% reduction in steps. Worth noting: an 80% cut in steps does not translate to an 80% cut in wall-clock time. Still, the official editing example completes in one pass at 2048×2048, and the text-to-image example supports portrait output up to 1680×2512.\n\nThe engineering details are considerate. The checkpoint ships with its recommended sampling schedule, which Diffusers loads automatically; setting `num_inference_steps` alone does not override the built-in schedule, and the default CFG is 1. In other words, developers get the officially calibrated 8-step configuration without manual tuning.\n\nCapability-wise, Turbo retains everything from 2.1: 2K image output, text-instruction editing, up to 10 reference images, and native transparent backgrounds. Ecosystem support carries over too — Diffusers, ComfyUI, vLLM-Omni, and SGLang have offered native support since the base release.\n\n## Runs Locally, But Read the License\n\nAt Alibaba Cloud Bailian's Beijing-region list price, Turbo costs 0.1 RMB per image and Pro costs 0.25 RMB — Turbo is 60% cheaper. If you would rather skip the API, the weights can be downloaded and run locally. But mind the license: the entire Qwen-Image-2.1 repository is under the Qwen Research License. Free use is limited to non-commercial research and evaluation; commercial development requires separate authorization. \"Open weights\" and \"use freely\" are two different things.\n\nCommunity feedback is mixed so far. Some users report clearly faster generation, while one tester on an RTX 5090 found that skin textures on generated people come out over-sharpened. There are not yet enough independent evaluations to confirm that Turbo's image quality matches the original — users who need stability may want to wait.\n\n## So What\n\nTurbo is more than an accelerated checkpoint. Competition in image generation is shifting from \"who draws best\" to \"who generates fastest and cheapest\" — once API pricing drops to about 0.1 RMB per image and local inference finishes in 8 steps, the economics of high-volume use cases like e-commerce batch imagery and draft assets change fundamentally. The Research License is also a reminder: when a model is announced as open-source, read the agreement before downloading the weights. As for how big the quality gap between 8 steps and 40 steps really is, let independent benchmarks answer that.\n\nReferences: [QwenLM\u002FQwen-Image-2.1 (GitHub)](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen-Image-2.1) · [Turbo weights (HuggingFace)](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen-Image-2.1-Turbo) · [NetEase report](https:\u002F\u002Fwww.163.com\u002Fdy\u002Farticle\u002FL8UQS3EP05566SCS.html)","qwen-image-2-1-turbo-8-step","2026-10-11T15:05:00Z","2026-10-11T15:10:46.237919Z",966,{"id":417,"title":418,"summary":419,"tags":420,"translations":429,"news_slug":435,"published_at":436,"created_at":437,"image_url":27,"view_count":438},"90606043-6747-4b31-b6a5-76d38c649d1c","Nano Banana 2.1:价格砍半,跑分压过Pro","Google 10 月 6 日将图像模型 Nano Banana 2.1 推到 GA,底座为 Gemini 3.6 Flash,价格较上代砍半:1K 图 3.36 美分、4K 7.56 美分。基准多角色一致性 1106 超过 Pro,但实测出图 Pro 仍占优。旧模型 10 月 29 日关停。",[421,424,427,428],{"id":422,"name":423,"slug":423,"description":27,"color":27},"a9524a82-a7c5-4daa-bb4b-a7ee77bb0b94","gemini",{"id":425,"name":426,"slug":426,"description":27,"color":27},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[430],{"id":431,"lang":39,"title":432,"summary":433,"content":434},"c05c0302-b428-41f1-a740-662505d6f7ff","Nano Banana 2.1: Half the Price, Beats Pro on Paper","Google's Nano Banana 2.1 hit GA Oct 6 on Gemini 3.6 Flash, halving image prices to 3.36 cents\u002F1K. Benchmarks top Pro; hands-on tests still favor Pro.","On October 6, Google pushed its image generation model Nano Banana 2.1 straight to general availability, skipping any preview. It replaces Nano Banana 2 (API name gemini-3.1-flash-image), released this February, with what Google calls across-the-board improvements: visual quality, prompt adherence, multi-turn character consistency, and text rendering all get better, and the base moves to Gemini 3.6 Flash — newer than the 3.1 that the Pro line still uses.\n\n## Prices cut in half first\n\nThe hardest change this time is price. A 1K image drops from 6.70 cents to 3.36 cents, a 4K from 15.10 cents to 7.56 cents — roughly half; the 2K tier sits at 5.04 cents. For comparison, Nano Banana Pro charges 13.40 cents per 1K image and 24.00 cents per 4K — a 4K image now costs about a third of Pro's price. Flash-tier speed is unchanged, with 1K\u002F2K\u002F4K resolutions supported (1K default).\n\n## Capabilities: 14 reference images and three thinking levels\n\nThe official documentation lists substantive updates: multi-image fusion now takes up to 14 reference images, with character consistency covering 4 characters and object fidelity covering 10 objects; tiling artifacts on ultra-wide aspect ratios (1:4, 1:8 panoramas) at 2K and 4K are fixed; infographic layout and text rendering accuracy are improved; and grounding with Google Web and Image Search is built in. Thinking comes in minimal\u002Fmedium\u002Fhigh levels (medium by default), with an input limit of 131,072 tokens.\n\n## Beats Pro in benchmarks, but hold your conclusions\n\nIn Google's model card benchmarks, 2.1 (with thinking) scores 1050 on overall preference, above Nano Banana 2's 990 and Pro's 935; infographic accuracy hits 0.521, nearly triple Gemini 3.1 Flash Image's 0.179; multi-character consistency reaches 1106, also beating Pro's 1011. But The Decoder's hands-on test offers a caveat: the previous NB2 also matched Pro in similar tests, while in practice Pro still delivers better color and proportions, and 2.1 stumbles on scale — drawing a horse more like a pony. Google's own positioning is merely Pro's \"more efficient counterpart.\"\n\n## So what\n\nThe old gemini-3.1-flash-image model shuts down on October 29, leaving existing workloads a three-week migration window. Google's playbook is now clear: use Flash-tier costs to turn image generation into default infrastructure — rolling out across the Gemini app, AI Mode in Search, AI Studio, Flow, Stitch, Google Ads, and Gemini Enterprise, with Pro reserved for premium budgets. When a 4K image costs 7.56 cents, the answer to \"what is an image worth\" is being rewritten: image generation is turning from a product selling point into a utility.\n\nReferences: [Google official model docs](https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs\u002Fmodels\u002Fgemini-nano-banana-2.1) · [The Decoder report](https:\u002F\u002Fthe-decoder.com\u002Fgoogles-new-image-model-nano-banana-2-1-generates-better-images-for-less-money\u002F)","nano-banana-2-1-halves-image-price","2026-10-08T07:10:00Z","2026-10-08T07:14:55.880447Z",462,{"id":440,"title":441,"summary":442,"tags":443,"translations":448,"news_slug":454,"published_at":455,"created_at":456,"image_url":27,"view_count":457},"473ca44d-b179-4c19-a7a6-916dd8b90e4a","FLUX 3 Image 发布:画框控图,逐框编辑","Black Forest Labs 发布 FLUX 3 家族图像组件 FLUX 3 Image:画布按 0-1000 网格,元素用边界框声明位置与内容,模型逐框生成;编辑只动目标框,其余像素不变。布局提示为全局说明加 JSON 元素表,可交由 LLM 规划。支持 10 张参考图、原生 4K 与商业权重自托管。",[444,445,446,447],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[449],{"id":450,"lang":39,"title":451,"summary":452,"content":453},"9b89cef8-30aa-4fd5-8706-15d5b8b1a9fd","FLUX 3 Image: bounding boxes, pixel-perfect edits, 4K","BFL's FLUX 3 Image: 0-1000 grid, caption plus JSON element tables, box-by-box edits leaving other pixels intact, 10 references, native 4K output.","Image generation's longest-running complaint is not quality but control: you write three lines of prompt, composition is a coin flip, and touching one corner means re-rolling the whole picture. On October 1, Black Forest Labs released FLUX 3 Image, the image component of its multimodal FLUX 3 family, and it makes control a first-class citizen — you draw the boxes, the model paints inside them, and every pixel outside stays put.\n\n## Bounding boxes as instructions\n\nFLUX 3 is BFL's multimodal family covering video, audio, images and actions; FLUX 3 Image handles image generation and editing. The core interaction is the bounding box. The canvas is a 0-to-1000 grid on both axes, and each element is declared with a box in [y_min, x_min, y_max, x_max] format plus a text description.\n\nA full layout prompt has two parts: a global caption describing the whole image, followed by a JSON element table where each row carries an id, a bounding box and a description. The caption cites every element by its id — animal_1 and friends — at first mention. This upgrades natural-language generation into structured declaration: position, content and relationships sit explicitly in the table, not guessed from one sentence.\n\nYou don't have to draw the boxes yourself. Hand an LLM one line and an aspect ratio, and it plans the caption and element table; every box stays editable afterward. On the model side, a prompt upsampler expands short requests into the dense captions FLUX 3 was trained on, but every box you drew reaches the model verbatim, with the same id and coordinates.\n\n## Edit one box, keep the rest\n\nEditing works box by box: re-describe a box, swap its content, or move it — everything untouched stays exactly where it was. BFL calls it pixel-perfect editing: add two divers, replace the duck, print on a shirt, round after round without the image falling apart. TechTimes frames this exact guarantee as the capability gap that has kept AI image editing unreliable in production workflows and LLM-driven pipelines.\n\nOther specs: up to 10 reference images per composition, native full-resolution rendering (an official example shows 5456 × 3072 output), and a commercial weights license that lets companies fine-tune and deploy on their own infrastructure.\n\n## An interface built for agents\n\nSeen as a feature update, this is nice; seen as interface design, it is clearly aimed at the agent era. JSON in, JSON out, verifiable coordinates, localized edits — exactly what an LLM pipeline needs when it consumes an image model: an upstream model plans the layout, a downstream model verifies box by box, and every step is checkable and reversible. Plain-prompt interfaces are black boxes by comparison; whether an edit landed correctly needs another vision model to guess.\n\nBFL says it outright on the page: an agent can use the model to create well-composed images with just a text prompt. Pair that with the commercial-weights licensing strategy and the target market is clearly bigger than designer tools — it is companies wiring image generation into automated workflows.\n\nThe takeaway: image generation is shifting from a wish-granting well to a plannable canvas. When position can be declared and edits localized, using image models starts to look like layout and engineering rather than gacha. Next time AI cannot edit your picture right, consider whether the problem is that you are still using one sentence — while the interface moved on to tables.\n\nReferences: official model page https:\u002F\u002Fbfl.ai\u002Fmodels\u002Fflux-3-image ; The Decoder https:\u002F\u002Fthe-decoder.com\u002Fblack-forest-labs-launches-flux-3-image-with-multi-step-editing-that-leaves-the-rest-of-your-picture-alone\u002F","flux-3-image-bounding-boxes","2026-10-02T23:06:31Z","2026-10-02T23:07:45.332994Z",1410,{"id":459,"title":460,"summary":461,"tags":462,"translations":467,"news_slug":473,"published_at":474,"created_at":475,"image_url":27,"view_count":476},"5757433d-e8ca-4371-89bc-f563dacc8089","程序能跑≠画对了:MaLiang框架专治AI作画跑偏","9 月 28 日 arXiv 论文 MaLiang-Harness 定义「程序-视觉」鸿沟:模型写的画图代码能跑通,构图、外观、运动却可能全错。框架以生成、检查、修订闭环推进,GPT-6-Astra 生成成功率 100%,96.0% 图像任务全指标达标,视频仅 76.9%;通用能力分预测不了视觉生成表现。",[463,464,465,466],{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},{"id":148,"name":149,"slug":149,"description":27,"color":27},[468],{"id":469,"lang":39,"title":470,"summary":471,"content":472},"c730f7b3-3874-43d2-bb1f-7f588e1c26e1","Code Runs, Image Wrong: MaLiang Tames the P2V Gap","arXiv paper MaLiang-Harness names the Program-to-Visual gap — code runs, image wrong — and closes it with a draft-inspect-revise loop for visual programs.","Letting models write code to draw has a pitfall everyone underestimates: the code runs, but the picture is wrong. Composition drifts, colors misfire, motion goes off track — the program is flawless at the syntax level and a mess at the visual level. A paper posted to arXiv on September 28 gives the phenomenon a name — the Program-to-Visual (P2V) gap — and ships a full construction-inspection-revision pipeline alongside it, reaching #2 on Hugging Face's daily paper ranking.\n\n## What the P2V gap is\n\nExecutable programs offer explicit control over how images and videos are constructed — that is their edge over prompts. But generating runnable code is only the beginning of visual creation: a program can execute without errors while violating the requested composition, appearance, or motion. The paper defines the discrepancy between \"code correct\" and \"picture acceptable\" as the P2V gap. The definition itself is useful: instead of vaguely saying a model \"draws poorly\", you get measurable slices — generation success, visual quality, and computational cost, tracked separately.\n\n## The fix: a draft-inspect-revise loop\n\nMaLiang-Harness has multimodal LLMs write drawing and animation programs, renders them to PNG images or silent MP4 videos through backends such as Canvas, SVG, and Three.js, then inspects the output and revises the code in a loop. Three mechanisms hold it together: PEG persists programs and task context, snapshotting every committed revision; TGP ties each edit to its rendered evidence; REV restores from historical revisions and re-checks the current one before export. Built on Deep Agents, the project ships a CLI, a chat interface, and batch evaluation tools, with code open-sourced on GitHub.\n\n## Results: 100% can generate, 76.9% actually pass\n\nThe team evaluated 11 closed-source MLLMs on their MaLiang-IBench (50 image tasks) and 4 models on MaLiang-VBench (13 video tasks). GPT-6-Astra hits 100% generation success on both benchmarks, with 96.0% of image tasks and 76.9% of video tasks passing all quality thresholds. The stinging part is the other finding: general capability scores mismatch visual-generation performance — models with similar overall scores differ substantially in satisfying visual requirements. Generic leaderboards lose nearly all predictive power here.\n\n## So what\n\nTwo takeaways. First, programmable generation is a real alternative to diffusion: when you need precise control over layout, structure, and motion, writing code is inherently more controllable than sampling — provided you close the self-check-and-revise loop, which MaLiang engineers into place. Second, stop trusting general benchmarks for model selection; composite abilities like visual generation demand task-specific benches. The project is named after Ma Liang — in Chinese context, the boy with the magic brush. The magic was never the brush; it was the eye that looks, and revises.\n\nReferences: [arXiv paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.34309), [GitHub repo](https:\u002F\u002Fgithub.com\u002Fgulucaptain\u002FMaLiang-Harness).","maliang-harness-p2v-gap","2026-09-30T15:20:00Z","2026-09-30T15:17:19.938024Z",970,{"id":478,"title":479,"summary":480,"tags":481,"translations":489,"news_slug":495,"published_at":496,"created_at":497,"image_url":27,"view_count":498},"88a287ca-c7c4-4f31-aff9-8a5116d02e5e","Qwen-Image-2.1:7B 轻量模型把生成+编辑焊进一套权重","阿里千问 9 月 20 日开源 Qwen-Image-2.1 图像模型,视觉生成仅 7B 参数,文生图与图像编辑合并为同一套权重,原生支持透明图像生成与编辑,最多支持 10 张参考图输入,官方基准反超 Nano Banana 2.0 等闭源对手。",[482,485,486,487,488],{"id":483,"name":484,"slug":484,"description":27,"color":27},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":250,"name":251,"slug":251,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[490],{"id":491,"lang":39,"title":492,"summary":493,"content":494},"f626c331-a8e6-4ace-a7b4-da6f6b83e6df","Qwen-Image-2.1: 7B Model Bakes Generation and Editing Into One Set of Weights","Alibaba's Qwen team open-sourced Qwen-Image-2.1 on September 20. The vision generation module weighs in at just 7B parameters, folding text-to-image and image editing into the same weights. It natively supports RGBA transparent image generation and editing, takes up to 10 reference images as input, and its official benchmark edged out closed-source peers such as Google's Nano Banana 2.0.","On September 20, Alibaba's Qwen team open-sourced Qwen-Image-2.1 — a diffusion model that collapses text-to-image and image editing into a single set of weights. The vision generation component is only 7B parameters, with weights mirrored to GitHub, Hugging Face, and ModelScope. A consumer GPU with 16GB or more of VRAM is enough to run it locally.\n\n## One checkpoint covers both generation and editing\n\nThe headline capability is native transparent (RGBA) image generation and editing. Most text-to-image models only emit RGB, forcing designers and e-commerce operators to run an extra matting pass to drop a subject onto a poster or product page. Qwen-Image-2.1 bakes the alpha channel into the training objective: prompts can request a layer with transparency built in, and edits preserve the transparent background while swapping expressions, replacing in-layer text, and so on. The transparent-image capability is inherited from the team's earlier specialised Qwen-Image-Layered model.\n\nThe same checkpoint also folds multi-reference conditioning and local editing into its input stage. Up to 10 reference images can be supplied, and three explicit control modes — marquee selection, brush strokes, and explicit masks — drive local edits. Across multiple rounds, facial details and product textures (logos, fabric patterns) stay consistent, which is the property e-commerce and portrait workflows actually need.\n\n## Stack: Single-Stream DiT with mixed-granularity attention\n\nThe architecture is a 32-layer Single-Stream Diffusion Transformer. Text and image tokens flow through one unified Transformer stream. System prefixes and editing instructions are routed through a token-level causal mask, while image patches use a chunk-level mask — two granularities running in parallel. Reference images and editing instructions are treated as static context: their Key-Value pairs are precomputed on the first step and reused on every subsequent step, which the team credits for the visible efficiency gains in multi-reference scenarios.\n\n## Benchmarks and where it sits in the ecosystem\n\nOn the team's own benchmark combining text-to-image and editing, Qwen-Image-2.1 edged out closed-source rivals such as Google's Nano Banana 2.0 and was listed by multiple outlets as the current open-source leader. The combination of fully open weights, unified generation-plus-editing, and native RGBA output is something Google's comparable products still don't offer.\n\nFor design, e-commerce, and content creation workflows, Qwen-Image-2.1's significance is not that it draws better pictures — it is that it removes the matting step and the model-switching step. One checkpoint handles generation and editing, and the native alpha channel pushes the workflow closer to a production-ready pipeline. When the open-source stack matches closed APIs on this kind of combined generation-plus-post capability, the marginal cost of design assets gets cut another notch.","qwen-image-2-1-open-source-7b","2026-09-20T16:00:00Z","2026-09-25T07:04:46.109836Z",1586,{"id":500,"title":501,"summary":502,"tags":503,"translations":508,"news_slug":514,"published_at":515,"created_at":516,"image_url":27,"view_count":517},"aa9b279e-07c6-4d0c-86f0-403c4af321fa","蚂蚁 RULER:六维评分表给 SVG 生成重写奖励信号","SVG 生成没有标准答案,CLIP 等标量指标用在矢量内容上会失真,当 RL 奖励还会诱发作弊。蚂蚁团队提出 RULER:每条指令转成六项评分细则,判分模型逐项打分加权成 GRPO 奖励,不需要配对真值和人工偏好。MMSVG 两基准评分从 0.432\u002F0.395 升到 0.693\u002F0.683。",[504,505,506,507],{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":167,"name":168,"slug":168,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[509],{"id":510,"lang":39,"title":511,"summary":512,"content":513},"7e37f23d-605b-4569-9422-cf2fd22cfb2f","Ant Group RULER: Rubric Rewards for SVG Generation","Ant Group's RULER replaces scalar SVG rewards with six-item rubrics. MMSVG scores rose from 0.432\u002F0.395 to 0.693\u002F0.683, matching DeepSeek-V3.","Getting language models to write SVG code is stuck at a fundamental point: the task has no ground truth. An assignment like \"draw an eye icon with eyelashes\" admits countless valid code solutions, with no paired reference images and no human preference labels. So both evaluation and training borrow scalar metrics calibrated on natural images — CLIP scores, aesthetic scores — but stylized vector content distorts them. The paper offers a vivid example: a faithful blue camera icon versus a visibly broken SVG get CLIP scores of 0.18 and 0.23 (the broken one scores higher), and aesthetic scores of 4.93 vs 4.84 are nearly indistinguishable, while a hand-designed rubric correctly assigns 0.94 vs 0.30. Worse, plugging such distorted metrics directly in as reinforcement learning rewards triggers reward hacking: the model learns to please the scorer instead of making the image good.\n\n## Ant Group's Fix: Swap Ground Truth for a Scoring Rubric\n\nA team from Ant Group with HKUST (Guangzhou) and Oxford researchers proposes RULER. The idea: dynamically generate a six-item scoring rubric for each generation instruction, spanning semantic, visual, and rendering-style dimensions — two items each. Semantic fidelity covers concept readability, major components, and prompt-specific relations; visual quality covers silhouette, form, and composition; rendering style covers finish and stylistic coherence. The rubric is derived from the text instruction alone by a frontier model (Claude-Opus-4.6); during training a judge VLM (Qwen3-VL-8B) scores rendered outputs item by item, and the weighted sum forms a fine-grained reward optimized via GRPO over the policy model (Qwen3-8B). The whole pipeline depends on no SVG ground-truth references and no human preference annotation.\n\n## Results: A Small Model Matches the Much Larger DeepSeek-V3\n\nOn MMSVG-Illustration and MMSVG-Icon, RULER lifts the rubric score from 0.432\u002F0.395 to 0.693\u002F0.683, surpassing dedicated SVG specialists and matching the substantially larger DeepSeek-V3. In blinded human preference over 150 prompts, non-tie win rates clear 50% against every baseline: 89.7% over the Qwen3-8B baseline, 66.1% over Qwen3-32B, 53.3% over VectorFusion, 66.9% over OmniSVG, and 96.5% over JanusCoder. Judge alignment with humans also opens a wide gap: Spearman correlation of 0.7929 (vs 0.6051 aesthetic, 0.5518 CLIP) and Goodman-Kruskal gamma pairwise ranking agreement of 0.7574. Ablations show that removing any rubric axis drops scores — all six items earn their place — and results hold when swapping in GPT-5.5 as rubric generator or running 4B\u002F8B base models.\n\n## So What\n\nThe point worth remembering is not about SVG per se, but about how to construct RL rewards for open-ended generation tasks: when a task lacks ground truth and scalar metrics distort, LLM-generated instance-aware rubrics offer a human-annotation-free path. Teams working on image generation, UI code generation, or creative writing — anywhere \"good answers are not unique\" — can port this instruction-to-rubric-to-itemized-judging-to-GRPO recipe directly. One caveat: rubric quality and judge-VLM capability become the new implicit ceiling, and the paper's ablations identify rubric design as the active lever for RL — the reward is only as reliable as the scorer.\n\nReference: [arXiv:2609.25270](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.25270) · Project page hangyuran.github.io\u002FRULER","ruler-rubric-rewards-svg-generation","2026-09-23T23:06:24Z","2026-09-23T23:09:18.077986Z",996,{"id":519,"title":520,"summary":521,"tags":522,"translations":527,"news_slug":533,"published_at":534,"created_at":535,"image_url":27,"view_count":536},"42d37ae0-3268-4684-91b1-9fca91f4e9c1","OpenAI 发布 ChatGPT Images 2.5:画个涂鸦就能出图,生成延迟砍半","OpenAI 周二发布图像模型 ChatGPT Images 2.5,生成延迟较 2.0 版最多降 50%,多轮编辑一致性更强;新增 Sketch 涂鸦功能,输入 @Sketch 画草图即可当参考图出图,配套图上圈注评论与 Templates 模板。已面向 ChatGPT、Work、Codex 用户全端推送。",[523,524,525,526],{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":328,"name":329,"slug":329,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[528],{"id":529,"lang":39,"title":530,"summary":531,"content":532},"377d26bc-be56-4e5d-86ee-f77c5c6d92dd","OpenAI launches ChatGPT Images 2.5: doodle-to-image Sketch, up to 50% lower latency","OpenAI's Images 2.5 cuts image generation latency by up to 50% versus Images 2.0 and adds Sketch: type @Sketch, doodle, and use your drawing as the reference for the final image. Comment-based editing and templates ship alongside. Rolling out to ChatGPT, Work, and Codex users on all platforms.","On September 8, OpenAI rolled out Images 2.5, the new version of its image generation model built into ChatGPT. The version bump looks incremental, but the more interesting shift is in the interaction: you can now doodle right inside the chat box and have the AI generate an image from your sketch.\n\n## From typing to drawing: what Sketch changes\n\nOpenAI's [official announcement](https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-chatgpt-images-2-5\u002F) puts it plainly: sometimes the clearest way to explain an idea is by drawing it. The new Sketch feature follows that logic — type @Sketch in ChatGPT, a canvas pops up, and you can sketch a room layout, the contour of an outfit, or just a stick figure, then add a short style description. The model turns the rough drawing into a complete image.\n\nA Verge editor tested it by drawing a rather rough cat with a computer mouse; the model still produced a realistic photo per the instructions, and kept refining details based on comments ([The Verge](https:\u002F\u002Fwww.theverge.com\u002Fai-artificial-intelligence\u002F991727\u002Fopenai-chatgpt-images-2-5-sketch)). The point: a composition like \"sofa on the left, floor-to-ceiling window on the right\" used to take rounds of textual prompt-tuning; now you can just draw the spatial relationship and let words handle style and details.\n\nTwo companion features shipped alongside: Templates, which provide starting frameworks for popular formats like flyers and product photos; and comment-based editing, which lets the model change only the marked parts of an image instead of rewriting the whole prompt.\n\n## The model itself: faster and steadier\n\nPer OpenAI's official claims, Images 2.5 cuts image generation latency by up to 50% compared with Images 2.0, produces more natural lighting and richer textures, better preserves the likeness of people, pets and objects from reference photos, and keeps details more consistent across multi-turn edits. TechRepublic and PCMag both confirmed the numbers, with TechRepublic noting that Axios — which had early access — tested it by designing a tattoo and iterating a logo through several rounds of notes without losing the original concept.\n\nImages 2.5 is available now to ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web, with no extra signup required.\n\n## Commentary: the parameter race is giving way to an interaction race\n\nMy take: Sketch matters more than Images 2.5 itself. Halved latency and better textures are the routine arms race every image model is running; \"doodle as prompt\" lowers the barrier from \"can you describe it\" to \"can you scribble two lines\" — a dimensional drop at the interaction layer.\n\nComment-based editing essentially imports the annotation workflow from design collaboration into the chat box. Multi-turn consistency plus localized edits add up to a complete loop for conversational image editing — a far friendlier path for non-professionals than learning a full prompt-engineering toolkit.\n\nA dose of realism: doodle input is not an OpenAI invention — sketch-to-image has been explored in the research community for years. What OpenAI did is productize it at the entry point used by hundreds of millions of users; the engineering value outweighs the invention value.\n\nSo what? If you generate images regularly, spend two minutes trying @Sketch. If you build AI products, note the signal: multimodal input is turning from a model capability into a UI default, and the next round of competition is happening at the interaction layer.","openai-chatgpt-images-2-5-sketch","2026-09-09T19:30:00Z","2026-09-09T19:09:39.971179Z",1016,{"id":538,"title":539,"summary":540,"tags":541,"translations":546,"news_slug":552,"published_at":553,"created_at":554,"image_url":27,"view_count":555},"a7095e70-a6c5-46b6-a616-446825bce5cf","Editable Visual Design：腾讯混元论文让 AI 设计稿带真实图层，直接拖拽改版","arXiv 2609.04034 提出用编码智能体替代扩散模型直出设计图：VLM 负责规划与审美判断，图像模型只按需生成独立素材，交付带真实文字、语义图层的 HTML，可拖拽编辑、可导出 PPT，Apache 2.0 开源仓库已近 500 星。",[542,543,544,545],{"id":120,"name":121,"slug":121,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[547],{"id":548,"lang":39,"title":549,"summary":550,"content":551},"8951802d-96be-4c50-b698-cf2505671a5e","Editable Visual Design: layered HTML posters from a coding agent","Tencent Hunyuan paper turns prompts into layered, mouse-editable HTML designs via a coding agent. Open source, Apache 2.0, near 500 GitHub stars.","Generating a poster with a diffusion model takes seconds—but what you get is a flattened bitmap. Text is error-prone, and moving a logo or changing a price line means regenerating the whole image. A paper published on arXiv on Sep 3, \"Editable Visual Design\" (arXiv:2609.04034), takes this pain point head-on. The paper page is listed under the Tencent Hunyuan organization, with 12 authors led by Junyan Ye; it picked up 42 upvotes on the Hugging Face papers page within two days, and the GitHub repository has climbed to nearly 500 stars.\n\n## The dilemma: diffusion looks good but cannot be edited; pure code is controllable but lacks taste\n\nThe paper lays out the shortcomings of both existing routes. Diffusion base models such as GPT-Image-2 and Nano-Banana are visually expressive, but end-to-end generation inherently yields flattened bitmaps with error-prone text, precluding layer-wise post-editing. Conversely, having a coding agent write code to draw gives precise layout control and decoupled layers, yet suffers from a lack of global aesthetic intuition and the difficulty of coding complex visual assets. Beautiful and editable have always been an either-or choice.\n\n## The fix: the VLM as the creative brain, the image model demoted to an asset shop\n\nThe division of labor is the interesting part: the VLM does the thinking—requirement comprehension, task planning, and aesthetic judgment—while the image generation model only shows up on demand to synthesize standalone visual assets. The whole pipeline runs an \"imagine first, then act\" closed loop: the agent generates isolated assets, writes native HTML\u002FCSS, and iteratively refines the design against visual rendering feedback. A key constraint is \"visual prior, without pixel shipping\"—the image model provides art direction such as composition, hierarchy, color, and spatial relationships, but its reference pixels never enter the delivered design; typography and layout are rebuilt in HTML by the agent. The final artifact carries real text and semantic layers, and users can adjust layouts by simply dragging in a graphical interface.\n\n## Beyond generation: replayable design\n\nAnother detail worth noting is Agent Design Replay: the system faithfully records the creative and reasoning trajectory of the agent, replayable like the working process of a professional human designer. The repository comparison table lists this explicitly—image generation has no creation path, direct coding is only partially replayable, while this approach turns \"how the design was made step by step\" into part of the artifact. The repo also ships deterministic checks covering the canvas, fonts, layer contracts, rendering, and editor round trips.\n\n## Shipping: two Codex skills, plus editable PPT export\n\nOn the engineering side the delivery is two independent Codex skills: editable-design generates polished designs from a brief, producing HTML, PNG, a mouse editor, an animated layer breakdown, and Agent Design Replay; html-to-pptx converts clean designs into editable PowerPoint files with independently selectable elements. The README recommends GPT-5.6 with high reasoning effort or above for the strongest results. The gallery provides 14 prompts across six visual-design categories—from a tea-launch campaign poster and a skincare e-commerce hero to an academic poster for the paper \"Attention Is All You Need\". The repository is open source under Apache 2.0, and the authors note it is a community project, not an official OpenAI project.\n\n## So what\n\nThe point worth remembering is not \"prettier generated images\" but a demotion of the generative model inside the visual pipeline: it no longer delivers; it only supplies assets and art direction. The deliverable becomes a software artifact with layer contracts—versionable, composable, replayable. Once a design draft turns from a bitmap into a code artifact, revision, reuse, and collaboration can finally be engineered. For teams building design tools and agent products, this division-of-labor paradigm is more instructive than any single benchmark score.\n\nReferences: arxiv.org\u002Fabs\u002F2609.04034; github.com\u002Fyejy53\u002FEditable-Design","tencent-editable-visual-design-agent","2026-09-06T13:08:40Z","2026-09-06T13:09:01.930576Z",904,{"id":557,"title":558,"summary":559,"tags":560,"translations":565,"news_slug":571,"published_at":572,"created_at":573,"image_url":27,"view_count":574},"c480d2d0-9156-4aa7-826f-fba2f252b6b7","蚂蚁开源 LLaDA-Image:6B 参数生成编辑一体,Turbo 版 4 步出图","蚂蚁 inclusionAI 开源 6B 参数的统一图像生成与编辑模型 LLaDA-Image:单 checkpoint 同时支持文生图与指令编辑,Turbo 版蒸馏到 4 步出图,官方自报 Qwen-Image-Bench 英文 53.53、中文 53.38;训练代码暂未放出。",[561,562,563,564],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[566],{"id":567,"lang":39,"title":568,"summary":569,"content":570},"0f5034ed-ff79-4049-be59-343c236f7284","LLaDA-Image: Ant's 6B open model unifies generation and editing","Ant's inclusionAI open-sources LLaDA-Image, a 6B model that generates and edits images in one checkpoint; the distilled Turbo runs in 4 steps.","Ant Group's inclusionAI open-sourced LLaDA-Image today (September 4): a 6B-parameter unified image generation and editing model family. Base and Turbo checkpoints shipped together, with BF16 and FP8 builds on both Hugging Face and ModelScope, plus inference code in the GitHub repository.\n\n## One checkpoint for both generation and editing\n\nThe core selling point is structure rather than raw fidelity: a single checkpoint handles text-to-image generation and reference-image, instruction-guided editing, with no separate editing backbone. Officially listed capabilities span four modes: text-to-image, VQ-conditioned generation, reference-image editing, and Chinese-English text rendering. The showcase samples cover photorealistic images, text posters, and instruction editing.\n\nTwo technical choices stand out. First, the text backbone and the DiT are both diffusion models trained in a unified framework — extending Ant's LLaDA line of diffusion language models into vision; in VQ mode, LLaDA2 produces image VQ tokens that SigVQ embeds before diffusion. Second, training is staged: image-only pre-training and mid-training build the visual prior first, paired language supervision comes later, and joint generation-editing training finishes the recipe.\n\n## The Turbo variant generates in 4 steps\n\nThe Base model recommends 50 sampling steps (guidance 5.0); the Turbo variant, distilled with Twin-DMD, compresses this to 4 steps (guidance 1.0), with the team citing 2-4 steps for fast generation and editing. On scores, the official report lists 53.53 (English) and 53.38 (Chinese) overall on Qwen-Image-Bench, described in the README as state-of-the-art — a self-reported result; cross-model comparison details sit in the technical report.\n\n## \"Fully open\" is missing one piece\n\nThe paper is titled \"Building Strong Image Generators with Fully Open Training Recipes,\" yet the repository's opensource plan shows inference code and weights released while training code is still marked \"coming soon.\" In other words, the \"fully open training recipe\" currently covers weights and inference; reproducing training means waiting. The technical report is on arXiv (2609.03796), and its Hugging Face papers page picked up 84 upvotes within a day. The implementation uses Python 3.11, PyTorch 2.8, and Diffusers 0.39.0, and the team credits ByteDance's VeOmni project.\n\nFor local-deployment users, a 6B model with an FP8 build puts it within reach of consumer GPUs. For industry watchers, the LLaDA diffusion lineage stretching from language models into image generation is a stronger signal than any single release. The \"fully open\" loop closes only when the training code lands.\n\nReferences: github.com\u002FinclusionAI\u002FLLaDA-Image; huggingface.co\u002FinclusionAI\u002FLLaDA-Image; arXiv 2609.03796","ant-llada-image-open-generation-editing","2026-09-04T13:09:21Z","2026-09-04T13:09:33.623707Z",2826,{"id":576,"title":577,"summary":578,"tags":579,"translations":585,"news_slug":591,"published_at":592,"created_at":593,"image_url":27,"view_count":594},"b51af942-496b-4b58-95fb-37980d12a743","逆向工程发现:微软画图本地生成的 AI 图像,像素里埋着服务器下发的水印 GUID","安全研究员逆向 Windows 画图与照片应用发现:Copilot+ PC 上本地跑 Stable Diffusion 生成的图像,提示词仍会送到微软服务器审核,服务器返回的唯一 GUID 被作为隐形水印写进像素,并与 C2PA 清单中的软绑定值一致。",[580,583,584],{"id":581,"name":582,"slug":582,"description":27,"color":27},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":189,"name":190,"slug":190,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[586],{"id":587,"lang":39,"title":588,"summary":589,"content":590},"8ff5fcc0-f6b8-454f-81f6-e43982987633","Reverse Engineering Finds a Server-Issued Watermark GUID Buried in Paint's Locally Generated AI Images","Reverse engineering of Windows Paint and Photos shows that images generated locally by Stable Diffusion on Copilot+ PCs still have their prompts sent to a Microsoft moderation server, whose unique GUID is embedded as an invisible pixel watermark matching the C2PA soft-binding value.","Generate an image with Windows Paint on a Copilot+ PC, and the diffusion model's inference really does run on your local NPU — but reverse engineering by security researcher Xusheng Li shows the result is not a purely \"local\" artifact: buried deep in the pixels is a unique GUID issued by a Microsoft server at prompt-moderation time. Full analysis at the source (https:\u002F\u002Fxusheng.dev\u002Fposts\u002Freversing\u002Fmspaint_invisible_watermark\u002Fmain\u002F).\n\n## Local generation ≠ local completion\n\nPaint and the Photos app both ship local AI models: four .onnxe files totaling roughly 370MB, with the generation backbone mager.onnxe alone at 302MB. These model files are obfuscated with a layer of XOR — one of the keys is the string Microsoft_2023 — and decode into standard ONNX models.\n\nThe catch: even on the local generation path, the user's prompt is still sent first to Microsoft's moderation service (endpoint path \u002Fv1\u002Fpaint-cocreator\u002Fmoderate-prompt). The server returns a revised prompt, a promptGenerationId, a watermarkId, and other fields. After the image is generated on-device, the watermarkId is written into the pixels — only the model inference is \"local\"; prompt moderation and identifier issuance both happen in the cloud.\n\n## How the watermark gets into the pixels\n\nThe watermark is applied by the WmkWriteWatermark function in Watermarker.dll. It strictly requires the payload to be exactly 16 bytes — precisely the length of a GUID. What actually gets embedded is an 18-byte message: the magic byte 0x4c, the 16-byte GUID, plus a one-byte checksum, expanded into 144 bits, with every bit required to be placed at least three times; images must be at least 192 pixels in both dimensions or the encoder returns an insufficient-capacity error.\n\nThe encoder performs content-adaptive, block-domain quantized modifications. Testing with a synthetic 512×512 image, the author found 193,376 of 262,144 pixels changed — invisible to the eye, yet roughly three-quarters of the picture altered. The failure semantics reveal the design intent: in Paint, a watermarking failure turns the entire generation into an error rather than returning an unmarked image; the Photos app is laxer, logging the failure and still returning the image.\n\n## Two layers of provenance, one GUID\n\nWhen saving, Paint also calls an online signing service to attach C2PA Content Credentials. Parsing the PNG chunks of a real generated image, the author found an 18,979-byte caBX manifest in which the c2pa.soft-binding field names the algorithm com.microsoft.invismark.1, and its value is exactly the watermarkId returned by the moderation service. The pixel-level invisible watermark and the file-level signed manifest are two layers of the same provenance system.\n\nSave formats for AI-generated results are restricted to PNG, JPEG, GIF, and .paint (BMP cannot embed a C2PA manifest); across successive generations, the previous promptGenerationId is echoed back with the next request, chaining multiple turns together.\n\n## So what\n\nMicrosoft's support page discloses content filtering and C2PA credentials, but not that the manifest carries a GUID bound to prompt moderation, let alone that it is written into the pixels. This does not conflict with Article 50 of the EU AI Act (effective August 2, 2026) — the regulation requires a detectable, machine-readable mark on AI content, but not a per-generation identifier that can be linked back to a moderation session. For users the lesson is direct: in system-level AI tools, the phrase \"local generation\" deserves renewed scrutiny. For the industry, watermarks are sliding from \"was this content AI-generated\" toward \"who generated it, in which session\" — the former is transparency, the latter is traceability, and blurring the two quietly dilutes the right to know.","mspaint-invisible-watermark-guid","2026-08-26T13:00:00Z","2026-08-25T19:08:40.472495Z",319,{"id":596,"title":597,"summary":598,"tags":599,"translations":604,"news_slug":610,"published_at":611,"created_at":612,"image_url":27,"view_count":613},"5ff06769-2251-40fd-a672-f394a1f68965","十人合影谁是谁:腾讯混元 WithEveryone 给群像生成装上身份锚点","腾讯混元团队发布 WithEveryone 群像生成框架:每个参考身份注入寻址 token,先规划身份-布局再渲染,用布局接地 ID 损失监督目标人脸。人脸相似度从 GPT-Image-2 的 0.462 提至 0.499,伪影从 0.169 降至 0.055。开源版因基模许可在重训。",[600,601,602,603],{"id":97,"name":98,"slug":98,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[605],{"id":606,"lang":39,"title":607,"summary":608,"content":609},"815191e2-cd9c-49ff-b44d-2db5e9bc7584","WithEveryone: Tencent Hunyuan's Group Image Framework Grounds Identity to Layout","Tencent's Hunyuan team released WithEveryone, a framework for identity-preserving group image generation with up to ten reference identities. Each identity enters as an addressed token; the model predicts an identity-layout plan before rendering, supervised by a Layout-Grounded ID Loss. Face similarity rises from GPT-Image-2's 0.462 to 0.499 while copy-paste artifacts drop from 0.169 to 0.055. An open-source version is being retrained due to base-model licensing.","A ten-person group photo where AI must place everyone correctly and keep every face right, guided by five to ten reference shots — this has long been the most failure-prone scenario in identity-preserving generation. WithEveryone, a framework released on arXiv by Tencent's Hunyuan-affiliated team ([paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.20336)), takes a different tack: don't start drawing until you have planned who stands where.\n\n## Why Identity Drifts as Groups Grow\n\nThe paper opens with the core tension: identity-preserving image generation becomes increasingly unreliable when a scene must contain many specified people. The model has to do three things at once — retain each identity, bind every reference face to a distinct person, and place that person at a correct location. Meanwhile, training-time identity losses must establish correspondence among several noisy predicted faces. When any of these links slips, you get face-swapping mix-ups, or worse, reference faces copied straight into the output.\n\n## Four Components: Plan First, Render Second, Supervise the Right Face\n\nWithEveryone's design breaks into four steps:\n\n- **Addressed tokens**: each selected identity is injected as a dedicated token;\n- **Identity-layout planning**: the model first predicts a structured identity-layout plan, then renders that plan as a visual condition — a seating chart before the photo;\n- **Layout-Grounded ID Loss**: annotated face regions supervise the intended identity directly, sidestepping unstable embedding-based face matching;\n- **ID Representation Forcing**: a prediction is additionally trained for each identity before image synthesis.\n\nThe keyword here is grounding: identity stops being an abstract concept floating in the prompt and gets anchored to a concrete region in the layout.\n\n## The Numbers: Similarity 0.462 to 0.499, Artifacts Cut by About Two-Thirds\n\nOn an identity-disjoint benchmark, the paper reports WithEveryone achieving the highest target-context identity similarity: face similarity improves from GPT-Image-2's 0.462 to 0.499, while copy-paste artifacts fall from 0.169 to 0.055 — roughly a two-thirds reduction. On coverage, it spans 97.3% of requested identities with a duplicate rate of only 2.8%. The paper's conclusion: explicit identity-layout grounding lets identity-preserving generation scale to larger groups without relying on direct reference-face copying.\n\n## A Practical Catch: No Checkpoints Yet\n\nThe detail worth noticing sits in the [GitHub repository](https:\u002F\u002Fgithub.com\u002Fdoby-xu\u002FWithEveryone) release plan: the research version in the paper is built on a foundation model whose licensing terms do not allow the checkpoint to be released. To give the community an open alternative, the team is actively training a new version on a foundation model that supports open release; code and checkpoints will follow once ready. In other words, for now you can read the method but not run the weights — a symptom of how base-model licenses increasingly constrain research reproducibility.\n\nThe paper made Hugging Face's Daily Papers list on August 21 with 39 upvotes, and the GitHub repo has 43 stars. For teams building multi-person scene synthesis or virtual group-photo applications, this belongs on the watchlist; real adoption, though, waits on the retrained open version. When the weights actually land — and whether the open release trades away any performance — is the leap that separates this technique from paper to tool.","witheveryone-group-image-identity-grounding","2026-08-24T13:00:00Z","2026-08-23T19:07:46.319322Z",716,{"id":615,"title":616,"summary":617,"tags":618,"translations":622,"news_slug":628,"published_at":629,"created_at":630,"image_url":27,"view_count":631},"6082cd23-0eca-40e0-9315-67318dc818ee","NovelAI Diffusion V5 发布:规模翻倍、32 通道 VAE,单次生成整页漫画","图像生成服务商 NovelAI 发布 Diffusion V5:规模超 V4.5 两倍,训练耗 26.8 万 B200 GPU 小时,VAE 从 16 通道升级到 32 通道。新模型强化自然语言与日语提示,测试中 22 角色同屏,单次生成多面板漫画页,原生支持 alpha 透明通道。",[619,620,621],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[623],{"id":624,"lang":39,"title":625,"summary":626,"content":627},"eae5e44c-795e-4f89-98c1-1ea7f070ec15","NovelAI Diffusion V5: Twice the Size, 32-Channel VAE, Comic Pages","NovelAI ships Diffusion V5 (Curated and Full): over twice the size of V4.5, trained on 268,000 B200 GPU-hours, with a custom 32-channel VAE. Stronger natural-language and Japanese prompting, up to 22 characters per scene, full multi-panel comic generation, and native alpha transparency.","NovelAI, the long-running anime-focused image generation service, has shipped its next-generation model. Announced on August 21, NovelAI Diffusion V5 arrives in two flavors, V5 Curated and V5 Full. The headline numbers from the official announcement are direct: built on the NovelAI in-house architecture, the model is more than twice the size of V4.5 and was trained in-house on 268,000 B200 GPU-hours ([official release post](https:\u002F\u002Fjournal.novelai.net\u002Fimage-generation-novelai-diffusion-v5-is-here-c2df7c6b8d2d\u002F)).\n\n## The Core Architectural Change: A 32-Channel VAE\n\nThe most technically interesting upgrade sits in the compression stage: the custom-trained VAE expands from 16 channels in V4.5 to 32 channels. The VAE represents images in latent space, and doubling the channel count directly determines how much detail the model can \"see\". The team cites fine lines in intricately drawn eyes, jewelry, text, and accessories as elements that now render with noticeably higher accuracy — less of the familiar \"melting goop\". Combined with the larger model size, complex environments featuring architecture and foliage also come through with more detail and coherence.\n\n## Prompting: Continuing the Shift to Natural Language\n\nV4.5 already let users write prompts in natural language instead of tags; V5 pushes that much further — the official framing is an unprecedented level of understanding. Officially supported languages are English and Japanese, with Japanese backed by training on Japanese text transcriptions. Testers also found Chinese, German, Spanish, and Portuguese prompts working, though results may vary since those were not a training focus.\n\n## Full Comic Pages in One Generation\n\nCharacter control is the other centerpiece. In testing, up to 22 distinct characters appeared on screen at once. Character Positioning moves from small fixed grids to free placement on the canvas, with the model following the positions you set and reducing feature bleed between characters. Longer prompts are supported as well. The comic capability is the V5 differentiator: describe the layout in natural language and a single generation produces a fully paneled page — the old training data only covered 2koma strips. Text rendering supports multiple languages including English, Japanese, and Chinese, and the frontend now auto-prepares the \"Text:\" block when you quote text.\n\n## What Is Not Finished Yet\n\nV5 does not ship complete: inpainting is available only with V5 Full at launch, while the curated inpainting model is still training — V5 Curated temporarily reuses the V4.5 inpainting function. Precise Reference and Vibe Transfer are also absent from launch day. The team says additional features will roll out after release, as usual. Meanwhile, the new custom VAE brings native alpha transparency: prompting \"transparent background\", \"has alpha\", or \"alpha transparency\" yields transparent backgrounds or translucent effects. The Enhance function gains a new \"Max✨\" tier backed by a new upscaler that also works standalone. The UI got a total overhaul alongside the release, and subscription usage limits were adjusted.\n\n## Why It Matters\n\nPlaced in the broader image-generation market, V5 reads as a vertical player reinforcing its moat. General-purpose models keep getting better at natural language, but anime styling, multi-character control, and comic layout are niche demands that general models rarely address with this depth — and the 268,000 B200 GPU-hour budget signals that the compute bar for this niche keeps rising. For creators, the real question is not which model is \"strongest\", but which workflow your work lives in. If single-shot full-page comic generation holds up in practice, it changes not just output speed but how storyboard iteration works.","novelai-diffusion-v5-release","2026-08-22T13:10:00Z","2026-08-22T13:07:28.385768Z",6519,{"id":633,"title":634,"summary":635,"tags":636,"translations":640,"news_slug":646,"published_at":647,"created_at":648,"image_url":27,"view_count":649},"774de6ac-98e1-4343-a67a-bfdc72d377bb","INFORMS 实证:AI 广告真实投放胜过设计师,18 个月后仍领先","INFORMS 发布的实地研究显示,一家户外活动公司的真实 Instagram 投放中,active-learning 引擎生成的 AI 广告在点击率上胜过人类设计师的 campaign,18 个月后无重训复测仍以 3.38% 对 3.24% 领先,且方差更低。",[637,638,639],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":189,"name":190,"slug":190,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[641],{"id":642,"lang":39,"title":643,"summary":644,"content":645},"4089072b-8f8f-4554-91b5-f3ebd27ca04e","INFORMS Field Study: AI Ads Beat Human Designers in Live Campaign, Lead Holds After 18 Months","A field study released by INFORMS shows that in a live Instagram campaign for an outdoor-activities company, ads generated by an active-learning engine beat the company's human-designed campaign on click-through rate, and still led 3.38% to 3.24% in a retest 18 months later with no retraining — and with lower variance.","The advertising industry has long held a comfortable belief about generative AI: the visuals are striking, sure, but when real money is on the line, human designers still win on brand discipline and taste. A field study released by INFORMS, the Institute for Operations Research and the Management Sciences, delivers a data point against that belief: in a live Instagram campaign, an AI-generated ad portfolio beat the human-designed one on click-through rate — and the advantage was still there 18 months later.\n\n## How the study was run\n\nThe researchers skipped lab simulations and plugged directly into a real campaign for an outdoor-activities company. They deployed an active-learning engine to generate background images promoting nature exploration, learning online as the campaign ran, with the twin goals of surfacing high-performing AI images while keeping them on-brand and effective. The comparison group was not a historical archive but the company's contemporary, professionally designed human campaign — which makes the contest much closer to \"AI system vs design team\" in real conditions.\n\n## The 18-month retest\n\nThe more telling part came later. Eighteen months after the original run, during a high-stakes booking season and without any model retraining, the team brought the AI portfolio back and compared it against the company's then-current human-designed campaign. The AI portfolio achieved a mean click-through rate of 3.38% versus 3.24% for the human campaign, again with lower variance — meaning it wasn't just higher on average, but more consistent. In an industry where creative fatigue is the norm, an unrefreshed AI portfolio outperforming contemporary human work 18 months on is arguably more interesting than the initial win.\n\n## Read the headline with a discount\n\nTo be clear, 3.38% versus 3.24% is a 0.14-percentage-point gap. In absolute terms, this is far less dramatic than the \"Dramatically Outperform\" phrasing in the press release headline — press releases are written for distribution, and judgment should rest on the numbers. The other caveats matter too: this is a single company, a single platform, and a single category, so generalizing to the entire ad industry deserves caution. And note what actually did the winning: the engine's ability to learn online from live campaign data, not the static quality of any single generated image. What likely won was not \"AI taste\" but the speed of the feedback loop — a human design iteration takes days, an active-learning iteration takes hours, and that compounding edge eventually shows up in CTR.\n\n## What it means for generative models in production\n\nFor the application side of image-generation models, the study carries two implications. First, the commercial value of generative models may lie less in \"drawing one picture for a designer\" and more in automated creative optimization once the model is coupled with a live feedback loop — high-iteration, low-per-image-value formats like Instagram backgrounds are exactly where this logic thrives. Second, it offers a harder evaluation standard for AI-generated content quality: not whether it looks good in a blind human rating, but whether it converts under real traffic. Notably, a separate recent study by Taboola with Columbia University used a \"sibling ads\" pairing method — matched pairs of AI-generated and human-made ads from the same advertiser, same campaign, same day — and reached a compatible conclusion: AI creative can now stand toe-to-toe with human work. A slice of the advertising industry is quietly turning from a \"creative business\" into an \"optimization problem.\"\n\nOne study doesn't remake an industry, but its method is worth remembering: put AI and humans on the same battlefield, under the same traffic, on the same timeline, and let them fight for real. If you're still using \"AI images aren't polished enough\" as the reason not to experiment, consider this — in performance advertising, the judge has never been polish. It's the data. (Study details: https:\u002F\u002Fwww.informs.org\u002FNews-Room\u002FINFORMS-Releases\u002FNews-Releases\u002FAI-Generated-Ads-Dramatically-Outperform-Human-Designers-in-Live-Campaign)","informs-ai-ads-beat-human-designers-18-months","2026-08-22T14:00:00Z","2026-08-21T17:14:43.621021Z",467,{"id":651,"title":652,"summary":653,"tags":654,"translations":661,"news_slug":667,"published_at":668,"created_at":669,"image_url":27,"view_count":670},"619ad304-0d2a-4dba-b91e-19414d036746","Grok Imagine Image 2.0：文生图 Arena 双榜第二","xAI 在 8 月 7 日上线 Grok Imagine Image 2.0,作为 grok.com\u002Fimagine、iOS、Android 的 Quality Mode 全面开放,API 名 grok-imagine-image-2.0。新模型主打'能用'——精准遵循指令、设计师级的版式排版、跨代视觉一致性;同时引入 5 张参考图编辑、区域级 magic wand、智能分割、背景移除、Smart resize 等编辑能力。在 Arena 图像编辑榜与文生图榜双榜第二,仅次于 OpenAI gpt-image-2。",[655,656,657,660],{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":658,"name":659,"slug":659,"description":27,"color":27},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",{"id":35,"name":5,"slug":5,"description":27,"color":27},[662],{"id":663,"lang":39,"title":664,"summary":665,"content":666},"5908e484-8bdb-4dd7-86d3-f3f41d1a0324","Grok Imagine Image 2.0: #2 on both Arena boards","xAI shipped Grok Imagine Image 2.0 on August 7, 2026, rolling it out as the Quality Mode on grok.com\u002Fimagine, iOS, and Android, with the API as grok-imagine-image-2.0. The new model is built for real work: tight prompt fidelity, designer-grade typography and layout, and stable visual identity across generations. Editing is a first-class citizen — magic wand, segmentation, background removal, 5-reference editing, and Smart Resize. Image 2.0 ranks #2 worldwide on both the Image Edit Arena and the Text-to-Image Arena, behind only OpenAI gpt-image-2.","## August 7: xAI puts a production-grade image model on grok.com\n\nxAI officially launched **Grok Imagine Image 2.0** on August 7. The new model ships as the Quality Mode on [grok.com\u002Fimagine](https:\u002F\u002Fgrok.com\u002Fimagine) and the iOS \u002F Android Grok apps, with the API going live the same day under the model name  ([official post](https:\u002F\u002Fx.ai\u002Fnews\u002Fgrok-imagine-image-2)).\n\nUnlike the previous generation's pure showmanship, 2.0 is explicitly built to be **usable in real work**: tight prompt following, designer-grade layout and typography, and stable visual identity across generations. In other words, xAI is no longer selling \"looks good\" — it is selling \"looks the same, every time\".\n\n## Editing as a first-class capability\n\nThe official post gives editing its own section, instead of tucking it under generation:\n\n- **Magic wand + Segmentation**: edit exactly the region you point at, leave the rest untouched;\n- **Background removal**: export the subject with a transparent background, ready to drop into downstream designs;\n- **Multi-ref editing**: a single generation accepts up to **5 input images**, removing the need to manually composite in Photoshop;\n- **Smart Resize**: pick a target aspect ratio (1:1, 2:3, 9:16, 16:9, ...) and the model refills the frame rather than cropping.\n\nOn top of that, **Templates** package common workflows (Photo Edit, Product Color Change, Editorial Product Poster, Reimagine, Photo Collage, Mascot Maker, BG Removal & Change, E-Commerce Photos, UGC Photos, Professional Headshot, Icon Maker, Character Sprite, Props & UI Kit, Emoji Creator, Merch Maker) as ready-made input-to-output pipelines.\n\n## Cross-generation consistency: scaffolding the world for video\n\nxAI positions Image 2.0 as the scaffolding for Grok's video stack. In the \"Build a world for video\" section, the official demo shows the same character, multiple locations, and several props being generated from a single prompt set, and then the same character reappearing in snow, blizzard, and a glacier tunnel — with **consistent character, art style, and prop detail across all images**.\n\nThis directly attacks the most painful problem in current video generation: cross-shot consistency. Image 2.0's answer is to fix the world's visual assets at the static-image layer, then hand off to Grok Imagine Video. The same logic runs through ByteDance's Seedream 5.0 → Seedance 2.5 image-to-video production line.\n\n## The scoreboard: Arena #2 on both leaderboards\n\nxAI discloses that Image 2.0 ranks **#2 worldwide on both the Image Edit Arena and the Text-to-Image Arena**, behind only OpenAI gpt-image-2. xAI models are listed on Arena under the name \"SpaceXAI\" (figures as of August 7, 2026).\n\nThe third-party Neura Market report on August 7 corroborates the same ranking and adds detail on Smart Resize, multi-reference input, and region-level editing ([Neura Market](https:\u002F\u002Fwww.neura.market\u002Fnews\u002Fxai-grok-imagine-image-2-0-editing-tools-arena-rankings)).\n\n## What it means for the industry: from \"generate\" to \"produce\"\n\nRead against the second half of 2026, the signal is bigger than the model itself:\n\n1. **\"Usable\" is more expensive than \"pretty\"**. The official page keeps saying \"real work\" and \"real creative work\" — xAI is now targeting design \u002F e-commerce \u002F marketing teams, not just creative tinkerers.\n2. **Editing is the new battleground**. Qwen-Image-2.0, Seedream 5.0, and Grok Image 2.0 are all betting on \"multi-ref + region edit\". The next six months of differentiation will shift from \"does the image look right\" to \"can it be edited precisely\".\n3. **Scaffolding video is the hidden agenda**. As ByteDance Seedance 2.5 and JD JoyAI-Video-Edit push video into the 30-second era, whoever can lock in visual asset consistency first gets the leverage on the video side. Grok's Image 2.0 → Imagine Video chain is exactly the same product logic.\n\n## How to use it\n\n- **Consumers**: open [grok.com\u002Fimagine](https:\u002F\u002Fgrok.com\u002Fimagine) and switch to Quality Mode, or use the iOS \u002F Android Grok app;\n- **Developers**: call ; docs at [xAI developer docs](https:\u002F\u002Fdocs.x.ai\u002Fdevelopers\u002Fmodel-capabilities\u002Fimages\u002Fgeneration#quick-start);\n- **Playground**: [xAI Console](https:\u002F\u002Fconsole.x.ai\u002Fteam\u002Fdefault\u002Fimage?model=grok-imagine-image-2.0&flow=explore).\n\n**So what**: the text-to-image race is not settled yet, and \"how productively you can edit a single image\" is the metric that will separate the field over the next six months. With editing as a first-class capability and an Arena #2 on both leaderboards, xAI no longer wants to be the \"image skin\" of Grok — it wants a slot in the designer's daily toolchain. For domestic players (Qwen \u002F Seedream \u002F Zhipu), cross-generation consistency and templated workflows are probably the most urgent lessons to take into the second half of the year.","grok-imagine-image-2-0-arena-second","2026-08-13T02:00:00Z","2026-08-12T22:03:23.353273Z",1725,{"id":672,"title":673,"summary":674,"tags":675,"translations":682,"news_slug":688,"published_at":689,"created_at":690,"image_url":27,"view_count":691},"74464bd0-01e0-45a9-bc27-f94b1f5966a1","Qwen-Image-3.0 上线千问 AI 平台：0.18 元\u002F张起步、Pro 与 Standard 双档开放 API","阿里巴巴千问图像生成模型 Qwen-Image-3.0 正式上线千问 AI 平台,旗舰版 Pro 与标准版 Standard 同步开放 API,文生图 0.18 元\u002F张起步,采用 1k\u002F2k 分档计费;海外用户可通过 Qwen Cloud 接入。这是 Qwen-Image-3.0 在 7 月底技术博客发布后首次对外商业化定价。",[676,677,678,679],{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":250,"name":251,"slug":251,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},{"id":680,"name":681,"slug":681,"description":27,"color":27},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[683],{"id":684,"lang":39,"title":685,"summary":686,"content":687},"fd51d80b-0765-4520-b53b-750d4f4590fe","Qwen-Image-3.0 launches: from 0.18 yuan per image via API","Alibaba's Qwen-Image-3.0 has officially launched on the Qianwen AI platform, with the Pro and Standard variants opening API access simultaneously. Text-to-image starts at 0.18 yuan\u002Fimage with 1k\u002F2k tiered pricing; international users can reach it via Qwen Cloud. This is Qwen-Image-3.0's first commercial API pricing after its late-July technical-blog release.","## Qwen-Image-3.0 finally has a price tag: from blog release to API commercialization\n\nIn early August, Qwen took the Qwen-Image-3.0 it had released at the end of July and wired it into its own API platform — which means a top-tier open-source Chinese text-to-image model now has a **clear official commercial price**, rather than forcing developers to either pull weights from Hugging Face or route through third-party proxies.\n\n**How the API is priced:** Text-to-image starts at 0.18 yuan\u002Fimage. The Pro version steps up to 0.5 yuan\u002Fimage at 2k resolution (1k stays at 0.25 yuan\u002Fimage); the Standard version does not differentiate resolution — both 1k and 2k are 0.18 yuan\u002Fimage. Input images are billed separately at 0.02 yuan\u002Fimage, but pure text-to-image generates no input fee. International users go through Qwen Cloud, where the dollar-equivalent price roughly sits in the 30%–50% range of comparable international models (source: Qianwen AI platform pricing docs at platform.qianwenai.com\u002Fdocs\u002Fdeveloper-guides\u002Fgetting-started\u002Fpricing).\n\n**How 1k vs 2k is decided:** The official rule uses **output image pixel area** to split tiers — area ≤ 2,250,000 pixels counts as 1k; anything above is 2k. A 1024×1024 image is roughly 1.05M pixels, which lands in 1k; 2048×1024 is about 2.10M, still 1k; you only hit 2k by outputting 2048×2048 or larger. This is an engineer-friendly tiering scheme — no floating-point edge cases when you write billing code.\n\n**Why 0.18 yuan\u002Fimage is worth highlighting:** Side-by-side against comparable Chinese open-source text-to-image APIs — Zhipu CogView-4 \u002F Kuaishou Kling \u002F Midjourney v7 in mainland China — **the Standard tier's 0.18 yuan\u002Fimage basically sits under the well-known 0.2 yuan psychological line**. The Pro tier does reach 0.5 yuan at 2k output, but combined with the Batch API's 50% off, it lands at 0.25 yuan, still competitive.\n\n**How does this relate to the late-July blog release?** That blog post (on the official Qwen site) covered the model architecture upgrades — native 2k, improved image-text understanding, more stable Chinese and English text rendering. **This release converts the model from \"downloadable\" to \"billable\"**. The two events sit on different tracks — R&D and commercialization — but reading them together, Qwen-Image-3.0 is now the **only open-source Chinese image model that simultaneously offers top-tier quality, an official API, and a complete price sheet**.\n\n**So what:** Open-source text-to-image has been competing for a long time, and model quality stopped being the deciding factor a while ago. The real moat is the length of the **\"download → integrate into app → bill and go live\"** chain. Qwen-Image-3.0's combination of **open weights + official hosting + clear tiered pricing** gives small and mid-sized teams the lowest-cost integration path to productize. If you hesitated before because \"downloading it meant building my own inference and calculating my own per-image cost,\" it is worth reassessing now — especially the Standard tier, which can almost mindlessly replace a self-hosted small model.\n\nReference: Qianwen AI platform pricing docs — https:\u002F\u002Fplatform.qianwenai.com\u002Fdocs\u002Fdeveloper-guides\u002Fgetting-started\u002Fpricing","qwen-image-3-0-qianwen-ai-platform-pricing","2026-08-05T03:00:00Z","2026-08-05T10:03:39.553754Z",1136,{"id":693,"title":694,"summary":695,"tags":696,"translations":701,"news_slug":707,"published_at":708,"created_at":709,"image_url":27,"view_count":710},"7cc1b87c-fe06-495a-9c01-9516d0c16354","腾讯混元 HunyuanImage-3.0 全面开源：80B 总参 \u002F 13B 激活的自回归 MoE，把多模态理解和生图拉到同一框架","腾讯混元 HunyuanImage-3.0 已全面开源：用统一自回归框架同时做多模态理解和图像生成，包含 64 个专家、80B 总参、13B 单 token 激活，是迄今规模最大的开源图像生成 MoE 模型；arXiv 同步发布 Technical Report（v3 修订于 2026-06-26），GitHub 提供权重、推理代码、CoT 思考、Image-to-Image 与蒸馏检查点。",[697,698,699,700],{"id":483,"name":484,"slug":484,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[702],{"id":703,"lang":39,"title":704,"summary":705,"content":706},"f0124d04-c0f5-4b9e-abc0-8c1acf5fe528","HunyuanImage-3.0 open-sourced: 80B MoE unifies understanding and T2I","Tencent Hunyuan Image-3.0 is fully open-sourced: it uses a unified autoregressive framework for multimodal understanding and image generation simultaneously, with 64 experts, 80B total parameters, and 13B per-token activation — the largest open-source image-generation MoE model to date. The accompanying arXiv Technical Report (v3 revised 2026-06-26) ships alongside weights, inference code, CoT thinking, Image-to-Image, and a distilled checkpoint on GitHub.","## When a frontier lab swaps diffusion for autoregression in open-source T2I\n\nTencent HunyuanImage-3.0 has publicly released its weights and inference code on GitHub (Apache-2.0), and the accompanying arXiv Technical Report v3 was revised on 2026-06-26. This is not \"another DiT variant\" iteration: Hunyuan replaced the diffusion + DiT route that the open-source T2I community has defaulted to, with a native multimodal autoregressive framework that packs \"understanding\" and \"generation\" into the same model.\n\nPer arXiv abstract and the GitHub README models card:\n\n- Total parameters: roughly 80B (\"over 80 billion parameters in total\" per arXiv)\n- Per-token activation: roughly 13B (\"13 billion parameters activated per token during inference\")\n- Experts: 64 (GitHub README, Key Features section)\n\nCombined with the \"largest and most powerful open-source image generative model to date\" self-claim from the README, this release pushes the \"open-source T2I\" bar up a notch. Previous open-source efforts (Imagen\u002FSD-class) were mostly dense DiT — HunyuanImage-3.0 is the first ~80B-scale, MoE-native, autoregressive open T2I model.\n\n## Not just a paradigm swap — the training pipeline is the story\n\nThe arXiv Technical Report attributes HunyuanImage 3.0's success to five components: meticulous data curation, advanced architecture design, native Chain-of-Thoughts schema, progressive model pre-training, and aggressive model post-training. This stack mirrors the company's Hunyuan-A13B LLM route — porting the MoE LLM playbook (data filtering + progressive pre-training + RL post-training) to image generation. The model does reasoning before generating, automatically enriching sparse prompts — LLM-style CoT stitched directly into visual generation.\n\nOn evaluation, the GitHub README gives two non-single-leaderboard metrics:\n\n- GSB (Human Evaluation): 1000 T2I prompts + 1000+ single\u002Fmulti-image editing cases, single-run, no cherry-picking; 100+ professional evaluators.\n- SSAE (Machine Evaluation): a proprietary Structured Semantic Alignment Evaluation that uses an MLLM to extract 3500 key points across 12 categories and score generation-vs-keypoint alignment.\n\nThese numbers do not claim \"X% better than SD\u002FImagen\" on a single benchmark — they signal SOTA-adjacent performance across 12 fine-grained dimensions on both automatic and human evaluation.\n\n## Three checkpoints + 8-step distillation close the \"frontier model → deployable\" gap\n\nThe GitHub README Model Card lists three checkpoints:\n\n| Model | Total \u002F Active | Recommended VRAM | Capabilities |\n|---|---|---|---|\n| HunyuanImage-3.0 | 80B \u002F 13B | ≥ 3 × 80GB | T2I |\n| HunyuanImage-3.0-Instruct | 80B \u002F 13B | ≥ 8 × 80GB | T2I + Text-Image-to-Image + Prompt Self-Rewrite + CoT Think |\n| HunyuanImage-3.0-Instruct-Distil | 80B \u002F 13B | ≥ 8 × 80GB | Same as above + fewer sampling steps (8 recommended) |\n\nDistil shipped the same day as Instruct (2026-01-26), distilling 50 steps down to 8. Not desktop-friendly — 8×80GB VRAM is not consumer hardware — but for second-tier inference hosts and serving vendors it means per-token cost can plausibly drop once throughput scales.\n\nInstruct makes editing first-class: T2I, Text-Image-to-Image, Prompt Self-Rewrite, CoT Think. The README showcases five canonical scenarios: CoT structured reasoning, complex-scene T2I, precise element-preserving editing, style transfer with thematic enhancement, and multi-image fusion (up to 3 inputs).\n\n## Why this story is worth writing\n\nWe wrote earlier about why frontier LLM vendors are betting on native multimodal rather than \"build a strong LLM first, then bolt on a vision encoder\". HunyuanImage-3.0 gives the inverse answer: take the image-generation branch and rebuild it on the same native multimodal autoregressive framework — collapsing \"understand → reason → generate\" into one path, skipping the DiT + LLM-via-API\u002Fencoder patchwork.\n\nThree wins:\n\n- CoT lands inside vision directly: the model \"thinks first, draws second,\" giving materially better robustness on sparse prompts than pure DiT.\n- Single model = single optimization objective: no separate vision and language tuning; RL signals flow end-to-end through post-training.\n- MoE leaves headroom for future scale: the 80B\u002F13B ratio pins inference cost at \"LLM-level\" — identical architecture pattern to Hunyuan-A13B, so the same inference stack (vLLM support, etc.) can be reused for T2I.\n\nThe cost is honest: 8 × 80GB is the entry ticket. This is not a personal toy — it sits in a different league from open-source peers like Flux-dev, Sana, or the broader Diffusers ecosystem.\n\n## So what\n\nThe \"T2I inside a native multimodal autoregressive framework\" route has been the closed-source direction for the past six months: OpenAI rebuilt DALL-E into GPT-image, Anthropic stacked vision tool calls on Claude, Google's Gemini went full-modal native. Tencent Hunyuan just crossed the line — not as an API demo, but with weights + inference code fully open-sourced at performance comparable to those closed-source routes.\n\nFor teams building in-house multimodal platforms (RAG + T2I + reasoning, all-in-one), HunyuanImage-3.0 Instruct's unified T2I + editing + CoT entry point is the form factor worth piloting. For application developers, the Distil-8 \u002F native-50 split defines the \"research vs product deployment\" choice.\n\nThe real story to watch in the next six months: can this MoE + autoregressive framework be ported to video generation at Sora 2 \u002F Veo 3 class? When someone releases video-native MoE AR weights in Q4, we'll write the comparison.\n\nReferences:\n- GitHub: \u003Chttps:\u002F\u002Fgithub.com\u002FTencent-Hunyuan\u002FHunyuanImage-3.0>\n- arXiv Technical Report v3: \u003Chttps:\u002F\u002Farxiv.org\u002Fabs\u002F2509.23951>\n- Tencent official: \u003Chttps:\u002F\u002Fhunyuan.tencent.com\u002Fimage>","tencent-hunyuanimage-3-moe-autoregressive","2026-08-05T01:00:00Z","2026-08-04T20:05:07.850913Z",1741,{"id":712,"title":713,"summary":714,"tags":715,"translations":720,"news_slug":725,"published_at":726,"created_at":727,"image_url":27,"view_count":728},"095917eb-02ae-4fd2-a1cb-17d0805442ee","微软 Mage-Flow 用 4B 跑赢 32B：原生分辨率 + 三件套协同设计把生成编辑都塞回单卡","微软亚洲研究院 \u002F 微软团队今天放出 Mage-Flow,一个 4B 规模的紧凑文生图与指令编辑一体化栈,代码与权重全部开源(github.com\u002Fmicrosoft\u002FMage)。\n\n核心思路很反直觉:不靠堆参数,靠 tokenizer + backbone + 系统三层协同。Mage-VAE 用一步扩散式编解码 + anchor-latent 正则,重建质量看齐 FLUX.2-VAE,但每像素编\u002F解码 MACs 砍到 1\u002F12 与 1\u002F22,直接把 VAE 从高分辨率瓶颈里捞出来。原生分辨率 MMDiT(NR-MMDiT)用 rectified flow 在 Mage-VAE 潜空间训练,配合可变长 FlashAttention + 逐样本 2D RoPE 打包,以及栈级 CUDA 算子融合,端到端训练吞吐提升约 2.5 倍,CFG 的条件\u002F无条件两支只需一次打包前向。\n\n效果上,Mage-Flow 与更大模型正面刚:在标准生成\u002F编辑基准上与 Qwen-Image 20B、Z-Image 6B、FLUX.2 32B、FireRed-Image-Edit 20B 打成平手甚至略胜;Turbo 变体在单张 A100、1024² 分辨率下生图 0.59 秒、编辑 1.02 秒,峰值显存 ~18-20 GB,在所有对比系统里最低。原生分辨率打包还支持 512 到 2048 之间任意长宽比,包括 4:1 这种极端比例,一个 checkpoint 通吃。\n\nMage-Flow 同时给出 Base、RL-aligned 和 4 步 Turbo 三档,以及独立的 Mage-Flow-Edit 编辑模型,后者在同一个图文条件框架里做语义编辑、外观迁移、修复与结构感知输出。整套都用 Diffusion-NFT 做后训练,提示跟随、文字渲染、美学与编辑一致性都有可观提升;Turbo 走对抗感知蒸馏,4 步出图,延迟低到可以做交互式应用。\n\n值得讨论的是这条路的范式意义:文生图\u002F编辑模型从 20B+ 卷回 4B 不是简单蒸馏,而是把 tokenizer 算力、骨干网络容量和工程栈代价作为一个整体来算账——压缩一个 VAE、压一次 attention、再把 kernel 拼一拼,4B 就能追平 32B。对开发者意味着可在单张消费级甚至云上入门卡上跑出可部署版本;对开源生态意味着 Qwen-Image 之后,Microsoft 又补了一个真正开源可商用的紧凑基础模型。接下来的关键问题是:Mage-Flow 这种「紧凑 + 系统协同」路线能否继续向下吃到 1B-2B,以及它和 Qwen-Image、FLUX.2 的差距是否会随 RL 与数据迭代再次拉大。",[716,717,718,719],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[721],{"id":722,"lang":39,"title":723,"summary":724,"content":724},"895ff5f9-6812-4ac1-80ca-6210ebe5eb32","Microsoft Mage-Flow: 4B beats 32B, generation and editing on one GPU","Microsoft Research Asia \u002F Microsoft team released Mage-Flow today, a 4B compact text-to-image + instruction-editing unified stack, with code and weights fully open-sourced (github.com\u002Fmicrosoft\u002FMage). The core idea is counter-intuitive: don't stack parameters, stack tokenizer + backbone + system. Mage-VAE uses one-step diffusion-style encoding\u002Fdecoding + anchor-latent regularization, matching FLUX.2-VAE in reconstruction quality, but cutting per-pixel encoding\u002Fdecoding MACs to 1\u002F12 and 1\u002F22, directly rescuing VAE from the high-resolution bottleneck. The native-resolution MMDiT (NR-MMDiT) uses rectified flow trained in Mage-VAE latent space, paired with variable-length FlashAttention + per-sample 2D RoPE packing, and stack-level CUDA operator fusion — end-to-end training throughput improves ~2.5x, and the CFG conditional\u002Funconditional two branches only need one packed forward. On results, Mage-Flow faces off against larger models head-on: on standard generation\u002Fediting benchmarks it ties or slightly beats Qwen-Image 20B, Z-Image 6B, FLUX.2 32B, FireRed-Image-Edit 20B; the Turbo variant on a single A100, at 1024² resolution, generates images in 0.59s and edits in 1.02s, peak memory ~18–20 GB, the lowest among all compared systems. The native-resolution packing also supports any aspect ratio between 512 and 2048, including extreme 4:1 — a single checkpoint handles everything. Mage-Flow also ships three tiers — Base, RL-aligned, and 4-step Turbo — plus an independent Mage-Flow-Edit editing model, the latter doing semantic editing, appearance transfer, inpainting, and structure-aware output in the same image-text conditional framework. The whole pipeline uses Diffusion-NFT for post-training; prompt following, text rendering, aesthetics, and editing consistency all see significant gains; Turbo uses adversarial-aware distillation to deliver 4-step output, with latency low enough for interactive applications. What's worth discussing is the paradigm significance of this path: text-to-image \u002F editing models going from 20B+ back to 4B isn't simple distillation — it's accounting for the tokenizer's compute, the backbone's capacity, and the engineering stack's cost as a whole. Compress the VAE once, compress the attention once, then piece the kernels together, and 4B can catch up with 32B. For developers it means a deployable version can run on a single consumer-grade or cloud entry-level card; for the open-source ecosystem it means, after Qwen-Image, Microsoft has added another truly open-source, commercially-usable compact foundation model. The key question going forward: can the Mage-Flow \"compact + system co-design\" path continue to be eaten down to 1B–2B, and will the gap with Qwen-Image and FLUX.2 widen again with RL and data iteration.","microsoft-mage-flow-4b","2026-07-23T03:30:00Z","2026-07-23T02:06:11.284736Z",1694,{"id":730,"title":731,"summary":732,"tags":733,"translations":738,"news_slug":743,"published_at":744,"created_at":745,"image_url":27,"view_count":746},"37eb94e4-2155-416a-a8d9-bc5075541a27","Qwen-Image-3.0 发布:把文生图从「好看」推向「好用」","7月21日,阿里千问正式推出 Qwen-Image 系列第三代图像生成基础模型 Qwen-Image-3.0。如果说 1.0 关键词是「准」、2.0 是「准、多、齐、美、真」,3.0 的核心只有一个字——「实」,把图像生成从「好看」推向「好用」。\\n\\n「实」落在三个维度。内容丰实:输入上限从 2.0 的约 1k token 跃升到 4.5k token,可一次性生成九宫格信息图、试卷、报纸、嵌套 UI 等复杂版面;官方演示中一张 3×3 信息图需 3.7k token 才能完整描述,刷新文生图模型的指令长度上限。细节真实:支持 10px 小字精准渲染,毛发、肌肤逼近摄影级,还能以原画笔法修复破损古画。知识厚实:原生支持 12 国语言渲染,覆盖 100+ 艺术风格,并能基于世界知识生成带分类标注的学术配图。\\n\\n这个转身很关键。文生图赛道长期困在「美学炫技」与「实用产出」的鸿沟里:小字看不清、版面装不下、多语言排不对,这些「丑但致命」的细节正是设计、教育、电商场景的最大门槛。Qwen-Image-3.0 选择把可用性当作头号指标。\\n\\n不过本次发布未同步开源权重或公开 benchmark,定价也未公布。「实」的承诺仍需第三方实测验证——尤其 4.5k token 下的稳定性与小字在非拉丁字体上的退化幅度。Qwen Studio 与阿里云百炼已开通 API 邀测,能否把文生图真正推过「生产力临界点」,值得持续跟踪。",[734,735,736,737],{"id":483,"name":484,"slug":484,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":250,"name":251,"slug":251,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[739],{"id":740,"lang":39,"title":741,"summary":742,"content":742},"7f08f157-61a0-4354-ab79-750460223652","Qwen-Image-3.0: from pretty pictures to useful generation","On July 21, Alibaba Qwen officially launched Qwen-Image-3.0, the third generation of the Qwen-Image image-generation foundation model. If 1.0's keyword was \"accurate\" and 2.0 was \"accurate, versatile, complete, beautiful, real\", then 3.0 has only one core word — \"practical\", pushing image generation from \"pretty\" to \"useful\". \"Practical\" lands in three dimensions. **Rich content**: the input upper limit jumps from about 1k tokens in 2.0 to 4.5k tokens, allowing one-shot generation of nine-grid infographics, exam papers, newspapers, nested UIs and other complex layouts; an official demo's 3×3 infographic needs 3.7k tokens to fully describe, refreshing the instruction-length upper bound for text-to-image models. **Real details**: supports 10px small-text precise rendering, hair and skin approaching photo-level, and can repair damaged ancient paintings in original brush techniques. **Solid knowledge**: native support for 12 languages in rendering, covering 100+ artistic styles, and can generate academic illustrations with classification labels based on world knowledge. This turn is critical. Text-to-image has long been stuck in the gap between \"aesthetic showing off\" and \"practical output\": small text unreadable, layouts can't fit, multi-language typesetting broken — these \"ugly but fatal\" details are the biggest barrier to design, education, and e-commerce scenarios. Qwen-Image-3.0 chooses to put usability as the headline metric. However, this release didn't open-source weights or publish benchmarks, and pricing isn't public. The \"practical\" promise still needs third-party empirical verification — especially the stability under 4.5k tokens and the degradation of small text on non-Latin scripts. API invites are open on Qwen Studio and Alibaba Cloud Bailian; whether text-to-image can really be pushed past the \"productivity critical point\" is worth tracking.","qwen-image-3-0","2026-07-21T08:00:00Z","2026-07-21T08:04:39.908821Z",865,{"id":748,"title":749,"summary":750,"tags":751,"translations":756,"news_slug":761,"published_at":762,"created_at":763,"image_url":27,"view_count":764},"f1397080-206a-469f-846c-932a4b3ab8f9","京东开源 JoyAI-Image：统一多模态基础模型，把「理解-生成-编辑」拧成一个闭环","京东（jd-opensource）正式开源 JoyAI-Image——一款真正意义上的统一多模态基础模型。它用一个 8B 多模态大语言模型（MLLM）做理解中枢，搭配 16B 多模态扩散 Transformer（MMDiT）做生成引擎，把「图像理解、文生图、指令式图像编辑」三件事拧进同一个模型族。\n\n最值得关注的是它的「理解⇄生成」闭环设计：以往的多模态系统理解归理解、生成归生成；JoyAI-Image 把两边打通——更强的空间理解反哺生成质量与可控编辑，而生成出的新视角又能为空间推理提供证据。这种双向激励的范式在国内开源多模态里相当少见。\n\n技术细节上，它在长文本排版、多视角生成、几何感知的空间编辑（Object Move \u002F Object Rotation \u002F Camera Control）上做了专项优化，在 spatial reasoning 与 Qwen-Image-Edit、Nano Banana Pro 等同台对标。模型权重、Diffusers 集成、ComfyUI 工作流、HuggingFace Demo 一并释放，Apache 2.0 完全开源。\n\n7 月 17 日最新更新：JoyAI-Image-Edit 与 Edit-Plus 原生支持 ComfyUI，可直接拖入工作流运行，无需额外依赖。\n\n这条路径的真正意义在于：当行业还在争论「理解派」和「生成派」谁主导，京东用统一架构给出了答案——边界本来就不必分。短期内，这会给 Qwen-Image-Edit、Seedream 等开源图像模型压力；长期看，「理解⇄生成」的双向激励范式有可能成为下一代多模态模型的标配。",[752,753,754,755],{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[757],{"id":758,"lang":39,"title":759,"summary":760,"content":760},"434cdef7-9db5-4188-8d22-2637aa642756","JoyAI-Image: JD.com's unified understanding-generation-edit loop","JD.com (jd-opensource) officially open-sources JoyAI-Image — a true unified multimodal foundation model. It uses an 8B multimodal LLM (MLLM) as the understanding center, paired with a 16B multimodal diffusion Transformer (MMDiT) as the generation engine, fusing \"image understanding, text-to-image, instruction-based image editing\" into a single model family. The most noteworthy piece is its \"understanding ⇄ generation\" closed-loop design: in traditional multimodal systems, understanding is understanding and generation is generation; JoyAI-Image bridges the two — stronger spatial understanding feeds back into generation quality and controllable editing, while newly generated viewpoints provide evidence for spatial reasoning. This bidirectional-incentive paradigm is rare in domestic open-source multimodal work. On the technical side, it's specifically optimized for long-text typesetting, multi-view generation, geometry-aware spatial editing (Object Move \u002F Object Rotation \u002F Camera Control), and stands against Qwen-Image-Edit, Nano Banana Pro on spatial reasoning. Model weights, Diffusers integration, ComfyUI workflow, and HuggingFace Demo are all released, fully open-source under Apache 2.0. Latest update on July 17: JoyAI-Image-Edit and Edit-Plus natively support ComfyUI and can be dragged directly into the workflow to run, with no extra dependencies. The real meaning of this path: when the industry is still arguing over \"understanding camp\" vs \"generation camp\", JD.com gives the answer with a unified architecture — the boundary didn't have to be drawn in the first place. In the short term, this puts pressure on open-source image models like Qwen-Image-Edit and Seedream; in the long term, the bidirectional-incentive \"understanding ⇄ generation\" paradigm may become the standard for the next generation of multimodal models.","jd-joyai-image","2026-07-20T06:00:00Z","2026-07-20T06:06:49.626150Z",1184,{"id":766,"title":767,"summary":768,"tags":769,"translations":774,"news_slug":779,"published_at":780,"created_at":781,"image_url":27,"view_count":782},"747917d5-e65b-46dd-b0db-40dfa119cdd1","Reve 2.1 用 Layout-First 架构 + 4K 输出登顶 Arena #2：用不到头部 1\u002F10 算力做独立图像生成实验室","Reve 实验室 7 月 9 日发布 Reve 2.1 文生图模型,距离 2.0 仅一个月即完成关键迭代。新模型在 Arena 文本到图像榜以 1306 Elo 重夺全球第二,单项编辑榜位列第八,并继续保持\"最高 4K 独立模型\"的头衔。\n\n技术上看,Reve 2.1 延续其核心押注——\"图像即代码\"。模型先生成层级化 layout 规划,再按区域独立渲染,所有元素天然可寻址、可单独重绘。这一代升级把规划精度、prompt 理解、外文渲染三件事一起拉高:4K 原生 16MP 输出在密集场景、细小文字、多语种文字同框的可控性都达到 SOTA 水平。\n\n最值得关注的反向信号是算力曲线。Reve 团队明确披露,2.1 的总训练算力不到头部玩家(微软、谷歌、Meta 等大厂图像生成产品)的十分之一,却跑出了 Arena Top 2 的实测成绩。在跨国 AI 实验室普遍堆算力、用十亿级图像-文本对训练扩散 Transformer 的当下,Reve 用\"代码化表征+极致工程效率\"反超,说明 layout 规划这一中间表示的密度红利还有大量未被挖掘。这意味着 2026 下半年的图像生成竞争,正在从\"模型规模竞赛\"分裂出\"表征效率\"这一独立赛道。\n\n对独立实验室和中小团队而言,这可能比又一个大模型发布更具方法论价值:当表征本身成为可编程对象,扩散模型的\"像素端到端\"假设就不再是唯一的最优解,设计工具、Agent 生图、批量素材管线都会出现新的工程入口。",[770,771,772,773],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[775],{"id":776,"lang":39,"title":777,"summary":778,"content":27},"89291c7d-aa39-4d1b-8242-6981240c2bfe","Reve 2.1 takes Arena #2 on 1\u002F10th the compute of the leaders","Reve Lab released its Reve 2.1 text-to-image model on July 9, completing a key iteration just one month after 2.0. The new model recaptures the global #2 spot on the Arena text-to-image leaderboard with 1306 Elo, ranks #8 on the image-editing sub-leaderboard, and continues to hold the title of \"highest-resolution 4K independent model\". Technically, Reve 2.1 continues its core bet — \"images as code\". The model first generates a hierarchical layout plan, then renders each region independently, so all elements are naturally addressable and can be individually redrawn. This generation lifts planning precision, prompt understanding, and foreign-language rendering together: 4K native 16MP output achieves SOTA controllability in dense scenes, small text, and multi-language text in the same frame. The most noteworthy counter-signal is the compute curve. The Reve team explicitly disclosed that 2.1's total training compute is less than one-tenth of that of the leading players (image-generation products from Microsoft, Google, Meta and other big labs), yet still hits Arena Top 2 in real-world testing. At a time when multinational AI labs are universally piling on compute, training diffusion Transformers on billion-scale image-text pairs, Reve uses \"code-style representation + extreme engineering efficiency\" to come out on top — showing that the density dividend of layout planning as an intermediate representation still has a lot of untapped room. This means image-generation competition in the second half of 2026 is splitting from the \"model size race\" into an independent track of \"representational efficiency\". For independent labs and small-to-medium teams, this may carry more methodological value than yet another large-model release: when representation itself becomes a programmable object, the diffusion model's \"end-to-end pixel\" assumption is no longer the only optimal answer, and design tools, Agent image generation, and batch asset pipelines will all see new engineering entry points.","reve-2-1-layout-first","2026-07-16T02:14:00Z","2026-07-16T02:15:14.149764Z",671,{"id":784,"title":785,"summary":786,"tags":787,"translations":792,"news_slug":797,"published_at":798,"created_at":799,"image_url":27,"view_count":800},"a6119d74-007a-4692-bb4e-d85b562a9d66","字节 SpectraReward：自我奖励 T2I 干翻 30× 大模型","字节 Seed × 港大 × 北大在 arXiv 2607.11886 提出 SpectraReward,把「读回 prompt」作为 T2I RL 的零样本奖励;Self-SpectraReward 让 BAGEL 自我奖励,GenEval 89.5 反超 Qwen3-VL-235B-A22B;核心结论:reward-policy alignment 比 reward scale 更重要。",[788,789,790,791],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[793],{"id":794,"lang":39,"title":795,"summary":796,"content":27},"719cfbbe-f9b1-4a13-8dcd-529ec042768e","SpectraReward: self-rewarded T2I beats 30x larger models","ByteDance Seed × University of Hong Kong × Peking University propose SpectraReward in arXiv 2607.11886, using \"reading back the prompt\" as a zero-shot reward for T2I RL; Self-SpectraReward lets BAGEL reward itself, with GenEval 89.5 beating Qwen3-VL-235B-A22B; the core conclusion: reward-policy alignment matters more than reward scale.","bytedance-spectra-reward","2026-07-15T04:30:00Z","2026-07-15T04:15:45.790333Z",787,{"id":802,"title":803,"summary":804,"tags":805,"translations":810,"news_slug":815,"published_at":816,"created_at":817,"image_url":27,"view_count":818},"28d4382c-abca-4e32-9294-24835089367d","GEAR 把 VQ tokenizer 拉回训练循环:让自回归图像生成首次跑赢 LlamaGen-REPA","北大 + 腾讯混元联合团队开源 GEAR(Guided End-to-End AutoRegression for Image Synthesis),通过 dual read-out 架构让 VQ tokenizer 和 AR 生成器真正端到端联训,在 ImageNet 上把 gFID 收敛速度推到 LlamaGen-REPA 基线的约 10 倍,并把对齐成本从 tokenizer 端转嫁到 AR 端,为下一代开源视觉生成模型提供了改动极小、可直接落地的训练范式。",[806,807,808,809],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":97,"name":98,"slug":98,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[811],{"id":812,"lang":39,"title":813,"summary":814,"content":27},"1d6553cc-5835-41b8-b50a-d6769e7e5706","GEAR pulls VQ tokenizers back into the training loop","Peking University + Tencent Hunyuan's joint team open-sources GEAR (Guided End-to-End AutoRegression for Image Synthesis), using a dual read-out architecture to truly end-to-end co-train the VQ tokenizer and AR generator, pushing gFID convergence speed on ImageNet to about 10× the LlamaGen-REPA baseline, and shifting the alignment cost from the tokenizer side to the AR side, providing a minimally changed, directly drop-in training paradigm for the next generation of open-source visual generation models.","gear-vq-tokenizer-llamagen","2026-07-10T14:05:00Z","2026-07-10T14:08:01.036643Z",156,{"id":820,"title":821,"summary":822,"tags":823,"translations":830,"news_slug":835,"published_at":836,"created_at":837,"image_url":27,"view_count":838},"06110002-82fb-421e-9daf-5ab73ead7f27","FourTune：把扩散模型后训练压进 4-bit，W4A4G4 让 FLUX.1-dev 12B 内存砍半、吞吐翻倍","扩散模型后训练一直被显存和吞吐拖住后腿——12B 级别的 FLUX.1-dev 想要做定制化或强化学习微调,成本高得吓人。MIT 韩松领衔的 FourTune (arXiv:2607.05711) 给出干脆解法:端到端把权重、激活、梯度全部压到 4-bit (W4A4G4),再叠一个 LoRA + frozen 数值稳定器并存的三分支混合管线,外加块级量化和定制 fused kernel,硬是在原生 4-bit 计算下把训练跑稳。在 FLUX.1-dev 12B 上,显存占用砍掉 2.25×、端到端吞吐提升 2.27×,且在定制化、强化学习、蒸馏三类任务上追平全精度微调的质量——没有因为量化而掉点。这与 OrbitQuant、FAIR-Calib 等偏推理侧的扩散量化路线形成互补,FourTune 直接瞄准后训练这一成本最高的环节,把 12B 级扩散模型的定制化门槛进一步拉低,W4A4G4 范式也可平移到 Wan、CogVideoX 等视频扩散模型,把 4-bit 后训练从工程 trick 升级成系统级方案。",[824,825,826,829],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":827,"name":828,"slug":828,"description":27,"color":27},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",{"id":35,"name":5,"slug":5,"description":27,"color":27},[831],{"id":832,"lang":39,"title":833,"summary":834,"content":27},"475c5513-8f91-4c89-92e1-13fc97dd4939","FourTune: 4-bit diffusion post-training, double throughput","Diffusion-model post-training has long been dragged down by memory and throughput — 12B-level FLUX.1-dev wanting to do customization or RL fine-tuning has prohibitively high cost. FourTune (arXiv:2607.05711) led by MIT's Song Han gives a clean solution: end-to-end compressing weights, activations, and gradients all to 4-bit (W4A4G4), plus a LoRA + frozen numerical-stabilizer coexisting three-branch hybrid pipeline, with block-level quantization and custom fused kernels, hard-running training stably under native 4-bit compute. On FLUX.1-dev 12B, memory usage is cut 2.25×, end-to-end throughput improved 2.27×, and on customization, RL, and distillation tasks the quality matches full-precision fine-tuning — no point loss from quantization. This complements diffusion-quantization routes for the inference side like OrbitQuant and FAIR-Calib, while FourTune directly targets the highest-cost link of post-training, further lowering the customization threshold for 12B-level diffusion models. The W4A4G4 paradigm can also transfer to video diffusion models like Wan and CogVideoX, upgrading 4-bit post-training from an engineering trick to a system-level solution.","fourtune-4bit-flux-12b","2026-07-09T00:01:00Z","2026-07-09T00:10:19.041104Z",694,{"id":840,"title":841,"summary":842,"tags":843,"translations":848,"news_slug":853,"published_at":854,"created_at":855,"image_url":27,"view_count":856},"19566223-1b02-4e48-8c44-518694edb049","Meta Muse Image 落地：Superintelligence Labs 把多模态推理与图生能力拧成一股","2026 年 7 月 7 日,Meta 在官方 Newsroom 正式推出 Muse Image——这是 Meta 超级智能实验室(MSL)的首个原生图像生成模型,已集成进 Meta AI 助手,并为 Instagram Stories 带来 30 余种 AI 创作特效,WhatsApp 聊天中也已开放图像生成能力(首批有限地区上线)。\n\n与 4 月发布的推理模型 Muse Spark 配对,Muse Image 并非简单的\"按文字画图\":它会先做多步规划——先理清画面版面、再调用实时 Web 上下文,最后融合多张视觉参考图。这意味着用户上传自拍加旅行照合成明信片、把宠物送进名画里,甚至根据一段历史地标文字直接渲染出照片级的画面,模型都能端到端完成,而不是停留在风格滤镜的层面。\n\n技术细节有三个看点:其一,可读文本渲染,模型能直接生成内嵌清晰文字的信息图与说明图;其二,@-mention Instagram 账号,系统会拉取该账号公开照片参与合成;其三,用户可在生成图上直接圈画修改,模型基于完整对话记忆持续迭代,而不会让上下文断裂。\n\n商业化方面,Meta 计划先把 Muse Image 推给广告主(通过 Advantage+ creative),再扩展到 Facebook 和 Messenger;视频侧,Muse Video 已经在路上,延续同一套\"推理 + 生成\"的双模型思路。\n\n如果说 GPT-5.6、Claude Sonnet 5 把\"主力 LLM\"打到新的成本\u002F能力曲线,那 Muse Image 的意义在于:Meta 终于补齐了\"原生图生 + 多模态推理 + 巨型应用分发\"这一完整闭环——把模型能力直接灌进三十亿用户的聊天框。",[844,845,846,847],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[849],{"id":850,"lang":39,"title":851,"summary":852,"content":27},"4db17b1f-56aa-4cd7-aa74-bbd2936e9c5a","Meta Muse Image: multimodal reasoning meets image generation","On July 7, 2026, Meta officially launched Muse Image in its official Newsroom — the first native image generation model from Meta's Superintelligence Labs (MSL), already integrated into the Meta AI assistant, bringing 30+ AI creative effects to Instagram Stories, with image generation also opened in WhatsApp chat (first launch in limited regions). Paired with the April-released reasoning model Muse Spark, Muse Image isn't simply \"draw from text\": it does multi-step planning first — first figuring out the layout, then invoking real-time web context, and finally fusing multiple visual reference images. This means users can upload selfies plus travel photos to synthesize postcards, send their pets into famous paintings, or even directly render photo-grade scenes from a piece of historical landmark text — the model handles it end-to-end, rather than staying at the level of style filters. Three technical highlights: first, readable text rendering — the model can directly generate info-graphics and explanatory images with clear text embedded; second, @-mention Instagram accounts — the system pulls that account's public photos for synthesis; third, users can circle-edit directly on the generated image, with the model iterating based on complete conversation memory rather than breaking the context. On commercialization, Meta plans to push Muse Image to advertisers first (via Advantage+ creative), then expand to Facebook and Messenger; on the video side, Muse Video is already on the way, continuing the same \"reasoning + generation\" dual-model thinking. If GPT-5.6 and Claude Sonnet 5 hit a new cost\u002Fcapability curve for the \"main LLM\", Muse Image's significance is: Meta finally completes the closed loop of \"native image generation + multimodal reasoning + mega-app distribution\" — directly pouring model capability into the chat windows of three billion users.","meta-muse-image","2026-07-07T20:01:00Z","2026-07-07T20:06:41.350434Z",1014,{"id":858,"title":859,"summary":860,"tags":861,"translations":866,"news_slug":871,"published_at":872,"created_at":873,"image_url":27,"view_count":874},"593fc68b-74f3-4ad9-b669-ed42b5d5da7a","iRDM 把经典 MMD 重新点燃:ImageNet 单步生成刷 SOTA,90 H200 小时把 FLUX.2 [klein] 蒸馏成一步","EPFL 团队 Lan Feng、Wuyang Li 和 Alexandre Alahi 在 7 月 2 日放出的 arXiv 2607.02375v1，让一年前被宣判“无法训练生成器”的经典 MMD 目标函数，在 ImageNet 单步图像生成上重夺王座。\n\n这篇 Improved Representation Distribution Matching (iRDM) 把“分布匹配”这个已经被 FLUX、SDXL 用烂的范式往深里挖——两条设计轴（分布怎么比较、在哪种表征里比较）系统梳理后，作者得出三条反直觉结论：(1) 老牌 MMD 只要估计方式对了，重新具备可扩展性；(2) generated batch size 才是关键变量，最优值要超过 2048，远超常规 batch；(3) 任何单一表征都能被“刷分”，必须用一组平衡的 encoder 做对照，并提出 SW_r14——基于 14 个 encoder 的 Sliced-Wasserstein 评估指标，专门防作弊。\n\n最炸裂的是落地数字：iRDM 在 ImageNet 单步生成上刷出 SW_r14 1.30 的 SOTA；PickScore 人类偏好评测中，71.2% 的样本被用户判定优于此前最佳单步模型。方法应用到 FLUX.2 [klein] 4-step 版本，仅用 90 H200 GPU 小时，把它后训练成单步生成器——GenEval 从 0.794 提升到 0.826，PickScore 从 22.58 提升到 22.76，单步反超四步。\n\n我的判断：MMD 复兴说明深度学习史上多次出现的“老方法重新崛起”剧情——核心目标函数的潜力可能比新架构还重要；FLUX.2 [klein] 90 GPU 小时被压缩成单步，单步生成正式具备“上生产”的资格，实时交互、内容平台、广告创意流水线都会跟进；项目页 + SW_r14 都开源，社区复现门槛极低。\n\n下一波关键问题：iRDM 能否套到视频生成？SW_r14 在文生图之外是否仍然防作弊？这些答案可能决定 2026 下半年扩散模型蒸馏的走向。\n\n来源：arXiv:2607.02375",[862,863,864,865],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[867],{"id":868,"lang":39,"title":869,"summary":870,"content":27},"5d4125a0-627f-4e7d-9a16-b4e2a160a5c6","iRDM revives MMD: one-step ImageNet SOTA, 90 H200-hours","EPFL's Lan Feng, Wuyang Li, and Alexandre Alahi, in arXiv 2607.02375v1 released on July 2, let the classic MMD objective function — which was pronounced \"unable to train a generator\" just a year ago — reclaim the crown on ImageNet single-step image generation. This Improved Representation Distribution Matching (iRDM) digs deeper into the \"distribution matching\" paradigm that FLUX and SDXL have overused — after systematically sorting out two design axes (how to compare distributions, in which representation to compare), the authors arrive at three counter-intuitive conclusions: (1) the classic MMD, as long as the estimation method is right, regains scalability; (2) the generated batch size is the key variable, with the optimal value needing to exceed 2048, far beyond the conventional batch; (3) any single representation can be \"score-gamed\", so a balanced set of encoders must be used for comparison, and the authors propose SW_r14 — a 14-encoder Sliced-Wasserstein-based evaluation metric, specifically to prevent cheating. The most explosive are the deployment numbers: iRDM posts SW_r14 1.30 SOTA on ImageNet single-step generation; in PickScore human preference testing, 71.2% of samples are judged by users to be better than the previous best single-step model. Applying the method to FLUX.2 [klein] 4-step version, with only 90 H200 GPU-hours, post-trains it into a single-step generator — GenEval improves from 0.794 to 0.826, PickScore from 22.58 to 22.76, single-step overtaking four-step. My judgment: the MMD revival is yet another \"old method rising again\" in deep-learning history — the potential of core objective functions may be more important than new architectures; FLUX.2 [klein] 90 GPU-hours compressed to single-step means single-step generation is officially qualified for \"production\" — real-time interaction, content platforms, and ad creative pipelines will all follow; project page + SW_r14 are both open-sourced, with very low community reproduction threshold. The next wave of key questions: can iRDM be applied to video generation? Does SW_r14 still prevent cheating beyond text-to-image? These answers may determine the direction of diffusion model distillation in the second half of 2026. Source: arXiv:2607.02375.","irdm-mmd-single-step","2026-07-06T06:30:00Z","2026-07-05T22:08:09.658235Z",829,{"id":876,"title":877,"summary":878,"tags":879,"translations":884,"news_slug":889,"published_at":890,"created_at":891,"image_url":27,"view_count":892},"52e96e5b-ec3b-4552-b94a-93cc2702ab84","Flow-Map GRPO：为确定性「少步生图」打开强化学习大门","\n少步 Flow-Map 生成器——如一致性模型、sCM、MeanFlow——过去两年一直是图像扩散模型里最快的一档：它们直接学习噪声到数据的「长途运输映射」，把采样步数压到个位数。但确定性正是它们的阿喀琉斯之踵：GRPO、PPO 这类需要随机轨迹和良好似然比的在线 RL 后训练方法，长期以来无法直接套用。\n\n**Flow-Map GRPO**（arXiv:2607.00535，7 月 1 日）解决了这一卡点。它的核心机制是 **ASFMC（Anchored Stochastic Flow Map Composition）**：通过基于锚点的条件重采样注入随机性，同时完整保留原始 Flow-Map 的边缘概率路径。这样既不破坏少步生成的高效性，又让 GRPO 目标函数可以求梯度。作者还推导出同时适用于**单步**和**两步** Flow-Map 参数化的 GRPO 目标，并在基于 FLUX 后端的 MeanFlow 与 sCM 上验证，多项奖励\u002F感知\u002F任务级指标全部上涨。\n\n最有看点的是它的工程哲学：**无需重新训练**。Flow-Map GRPO 把后训练做成「外挂」模块，直接对预训练好的确定性生成器做对齐，不需要改参数化，也不必把模型再训成原生随机模型——对存量 checkpoint 极其友好，这也意味着 FLUX、Qwen-Image、Wan 等已经上线的少步工作流未来都可以低成本接 RL。\n\n对 GRPO+扩散社区，这是继 6 月末 Qwen-Image-2.0-RL 技术报告之后的又一个信号：**后训练范式正在从 LLM 辐射到图像生成**，而少步 Flow-Map 是这条扩展路径上一直缺位的一环。Flow-Map GRPO 补齐它以后，行业剩下的主要是工程问题——更稳定的奖励模型、更大规模的人类偏好对齐、面向特殊风格的可控强化，都会沿着这个口子铺开。",[880,881,882,883],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":189,"name":190,"slug":190,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[885],{"id":886,"lang":39,"title":887,"summary":888,"content":27},"fb27eadf-41c8-4507-ba32-f6800410047d","Flow-Map GRPO opens RL for deterministic few-step imaging","# Flow-Map GRPO: opens the door to reinforcement learning for deterministic \"few-step image generation\" Few-step Flow-Map generators — like consistency models, sCM, MeanFlow — have been the fastest tier in image diffusion models over the past two years: they directly learn the \"long-distance transport map\" from noise to data, compressing the sampling steps to single digits. But determinism is their Achilles' heel: online RL post-training methods like GRPO and PPO that need random trajectories and good likelihood ratios have long been unable to be directly applied. **Flow-Map GRPO** (arXiv:2607.00535, July 1) solves this bottleneck. Its core mechanism is **ASFMC (Anchored Stochastic Flow Map Composition)**: injects randomness through anchored conditional resampling while fully preserving the original Flow-Map's marginal probability path. This doesn't break the efficiency of few-step generation, while making the GRPO objective function differentiable. The authors also derive GRPO objectives applicable to both **one-step** and **two-step** Flow-Map parameterizations, and verify on MeanFlow and sCM based on the FLUX backend, with multiple reward \u002F perception \u002F task-level metrics all rising. The most noteworthy is its engineering philosophy: **no retraining needed**. Flow-Map GRPO makes post-training a \"plug-in\" module, directly aligning pretrained deterministic generators, without changing parameterization, and without having to retrain the model into a native stochastic model — extremely friendly to existing checkpoints, which also means FLUX, Qwen-Image, Wan and other already-online few-step workflows can in the future all hook up RL at low cost. For the GRPO+diffusion community, this is another signal after the late-June Qwen-Image-2.0-RL tech report: **the post-training paradigm is radiating from LLM to image generation**, and few-step Flow-Map has been the missing link on this extension path. After Flow-Map GRPO fills it, the remaining things for the industry are mainly engineering — more stable reward models, larger-scale human-preference alignment, controllable RL for specific styles, will all unfold along this opening.","flow-map-grpo-few-step","2026-07-05T12:02:00Z","2026-07-05T12:04:44.695763Z",677,{"id":894,"title":895,"summary":896,"tags":897,"translations":902,"news_slug":907,"published_at":908,"created_at":909,"image_url":27,"view_count":910},"e9688664-6ec6-4816-8898-3a92e6638c7a","MrFlow：四步分阶段采样把文生图扩散推到 10× 加速，OneIG 损失压到 1%","北航郑星宇等人在 arXiv 2607.01642 提出 MrFlow —— 一种完全 training-free 的多分辨率扩散加速方案。FLUX.1-dev、Qwen-Image 把文生图质量推到开源 SOTA，但每张图动辄要跑几十步 denoising，推理成本成为落地的主要门槛。timestep distillation 需要为每个底模重新训练；现有 training-free 多分辨率方案在潜空间上采样，又常出现明显模糊与伪影。MrFlow 把推理拆成四步显式流水线：低分辨率主体生成 → 像素空间 GAN 超分 → 低强度噪声注入 → 高分辨率细节精修，无需训练、无需运行时动态判别。在 FLUX.1-dev、Qwen-Image 上实现 10× 端到端加速，OneIG 损失控制在 1% 以内；与 timestep distillation 正交叠加可达 25×。这是半年内 diffusion 加速路线里少见的\"真正可落地\"方案 —— 不绑模型、不依赖定制 kernel，对开源社区尤其友好，暗示了\"加速可以一层一层叠加\"的新范式。代码已开源至 GitHub（Xingyu-Zheng\u002FMrFlow）。",[898,899,900,901],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[903],{"id":904,"lang":39,"title":905,"summary":906,"content":27},"9bed4383-ee98-4c28-ba59-b825c475c5ad","MrFlow: staged sampling gives 10x text-to-image speedup","Zheng Xingyu et al. at Beihang University, in arXiv 2607.01642, propose MrFlow — a completely training-free multi-resolution diffusion acceleration scheme. FLUX.1-dev and Qwen-Image have pushed text-to-image quality to open-source SOTA, but each image often requires dozens of denoising steps, making inference cost the main barrier to landing. Timestep distillation requires retraining for each base model; existing training-free multi-resolution schemes often exhibit obvious blur and artifacts when upsampling in latent space. MrFlow splits inference into an explicit four-step pipeline: low-resolution main body generation → pixel-space GAN super-resolution → low-strength noise injection → high-resolution detail refinement, no training, no runtime dynamic discrimination needed. On FLUX.1-dev and Qwen-Image, it achieves 10× end-to-end speedup with OneIG loss controlled within 1%; orthogonal stacking with timestep distillation can reach 25×. This is a rare \"truly landable\" solution in the past half year's diffusion acceleration routes — not tied to a model, not relying on custom kernels, especially friendly to the open-source community, hinting at a new paradigm of \"acceleration can be stacked layer by layer\". Code is open-sourced on GitHub (Xingyu-Zheng\u002FMrFlow).","mrflow-four-step-t2i","2026-07-04T14:11:00Z","2026-07-04T14:12:49.654750Z",752,{"id":912,"title":913,"summary":914,"tags":915,"translations":920,"news_slug":925,"published_at":926,"created_at":927,"image_url":27,"view_count":928},"22aad878-387f-4166-b68c-896b39a27de3","Sourceful Riverflow 2.5：把「评分函数」塞进图像生成，让「什么是好图」变成可编程的","## Sourceful Riverflow 2.5：把评分函数塞进图像生成，让什么是好图变成可编程的\n\n2026 年 6 月，曼彻斯特的 Sourceful 推出 Riverflow 2.5 Pro，主打 reasoning image generation。在 Image、Graphic Design、Image Editing 三个 Design Arena 维度同时登顶，把 GPT Image 2、FLUX.2、Midjourney V8.1 这些通用偏好选手挤到身后。和大多数图像模型「越像大众审美越好」的路线不同，Riverflow 2.5 走的是一条更工程化的路：把生成本身拆成多步推理 + 自定义评分 + 自审稿流程，让品牌方第一次能把「什么是合格图」用代码写出来。\n\n技术上，Riverflow 2.5 的关键创新不是更大的扩散 Transformer，而是一条显式的 reasoning 链路。用户可以指定 thinking level（low 到 xhigh 四档），系统会按这个预算决定「愿意重画几轮 + 评判多严格」。更核心的是 custom scoring rubric：同一个 prompt 和源图，只换评分准则，模型就能在「包装批准 92.3%」「付费社交流量 90.4%」「首页 hero 89.2%」三种目标之间自由切换。配合最多两份自定义字体文件的 Font Control、1K\u002F2K\u002F4K 输出，以及 Transparent \u002F Solid \u002F Normal 三种背景模式，它把「出图-审稿-交付」压成了一个 agent 可编排的接口。\n\n这件事的意义不止于一家英国创业公司多了一款产品。当 Claude Code、Codex、ChatGPT Agent 开始大批量调度图像模型时，模型的差异化就从「画面好不好看」转向「能不能被 agent 信任地调用」。Riverflow 2.5 押注的方向——把评判标准外置为可编程参数——正是 agent 时代视觉生成最稀缺的一块拼图。",[916,917,918,919],{"id":483,"name":484,"slug":484,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[921],{"id":922,"lang":39,"title":923,"summary":924,"content":924},"1846b557-3267-4053-adc1-913212a102fb","Sourceful Riverflow 2.5: putting scoring functions inside generation makes \"a good image\" programmable","In June 2026, Manchester-based Sourceful released Riverflow 2.5 Pro, built around reasoning image generation. It topped all three Design Arena dimensions — Image, Graphic Design, and Image Editing — pushing general-preference leaders like GPT Image 2, FLUX.2, and Midjourney V8.1 behind it. Unlike most image models, which chase mass aesthetics, Riverflow 2.5 takes a more engineering-driven path: it decomposes generation itself into multi-step reasoning, custom scoring, and a self-review pipeline, letting brands define \"what counts as an acceptable image\" in code for the first time.\n\nTechnically, the key innovation is not a bigger diffusion Transformer but an explicit reasoning chain. Users pick a thinking level (four tiers from low to xhigh), and the system uses that budget to decide how many redraw rounds to spend and how strict the judging gets. The core piece is the custom scoring rubric: with the same prompt and source image, swapping only the scoring criterion switches the model between goals like \"packaging approval 92.3%\", \"paid social traffic 90.4%\", and \"homepage hero 89.2%\". Combined with Font Control for up to two custom font files, 1K\u002F2K\u002F4K output, and Transparent \u002F Solid \u002F Normal background modes, it compresses generate-review-deliver into a single interface an agent can orchestrate.\n\nThe significance goes beyond one more product from a British startup. As Claude Code, Codex, and ChatGPT Agent start dispatching image models at scale, differentiation shifts from \"how good the picture looks\" to \"whether the model can be trusted by an agent\". Riverflow 2.5's bet — externalizing the judgment standard as a programmable parameter — is exactly the missing piece visual generation needs in the agent era.","sourceful-riverflow-2-5","2026-07-01T10:00:00Z","2026-07-01T16:10:36.171648Z",956,{"id":930,"title":931,"summary":932,"tags":933,"translations":938,"news_slug":943,"published_at":944,"created_at":945,"image_url":27,"view_count":946},"b20f76a0-85ec-4de4-b094-21582692ba53","Qwen-Image-2.0-RL 技术报告：把 GRPO+OPD 整套后训练范式搬进文生图扩散模型","Qwen 团队发布 Qwen-Image-2.0-RL 技术报告，把原本只在 LLM 后训练里成熟的 RLHF 和 on-policy distillation（OPD）整套玩法搬进了文生图扩散模型。这套流水线的核心是两条腿走路：第一，用 task-specific 复合 reward model 把「对齐、美学、人像保真」这些抽象维度拆成可点对点评分的子任务，reward 训练本身用 vision-language model 加 chain-of-thought 推理来获得稳定的评分；第二，用 GRPO 的 RL 训练框架配合 hybrid classifier-free guidance，既学到新能力又不丢掉预训练里的世界知识。Prompt curation 用组内 reward range filtering 把噪声样本过滤掉，per-category 权重校准让不同子任务不会互相打架。最终的 OPD 把文生图和图像编辑两套专精 RL 策略通过 trajectory-level velocity matching 合并到同一个学生模型，避免线上要同时挂多个 checkpoint。结果是 Qwen-Image-Bench 综合分从基座提升到 57.84（+2.61），T2I Arena Elo 涨 78 到 1193，编辑 Arena 涨 93 到 1349。这标志着文生图正在从「规模 + 数据」的预训练范式过渡到「RL + 蒸馏」的后训练范式——和 LLM 在 2024 年走过的路几乎是同构的。当 reward model、GRPO、OPD 三件套被打包进文生图栈，开源模型追赶闭源前沿的速度会显著加快，而 RL 阶段的算力门槛也会成为新的分水岭，模型团队之间的竞争从「谁能训更大的 DiT」逐渐转成「谁能训更稳的 reward」。",[934,935,936,937],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":250,"name":251,"slug":251,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[939],{"id":940,"lang":39,"title":941,"summary":942,"content":942},"ae50980d-9228-400c-b8dd-b8aecab986c4","Qwen-Image-2.0-RL technical report: bringing the full GRPO+OPD post-training stack into text-to-image diffusion","The Qwen team released the Qwen-Image-2.0-RL technical report, porting the RLHF and on-policy distillation (OPD) playbook — matured in LLM post-training — into text-to-image diffusion models. The pipeline stands on two legs. First, a task-specific composite reward model decomposes abstract dimensions like alignment, aesthetics, and portrait fidelity into subtasks that can be scored point by point; reward training itself uses a vision-language model with chain-of-thought reasoning to obtain stable scores. Second, GRPO serves as the RL training framework, paired with hybrid classifier-free guidance to learn new capabilities without losing the world knowledge from pretraining. Prompt curation uses within-group reward range filtering to drop noisy samples, and per-category weight calibration keeps different subtasks from fighting each other. The final OPD step merges the two specialized RL policies — text-to-image and image editing — into a single student model via trajectory-level velocity matching, so serving doesn't need to run multiple checkpoints online.\n\nResults: the Qwen-Image-Bench composite score rises to 57.84 over the base (+2.61), T2I Arena Elo climbs 78 to 1193, and the editing Arena gains 93 to reach 1349. This marks text-to-image shifting from the \"scale + data\" pretraining paradigm to an \"RL + distillation\" post-training paradigm — nearly isomorphic to the path LLMs took in 2024. Once the reward model, GRPO, and OPD trio ships inside the text-to-image stack, open-source models will close in on the closed frontier much faster, and the compute bar of the RL stage becomes the new dividing line: competition moves from \"who can train the bigger DiT\" to \"who can train the more stable reward\".","qwen-image-2-rl-grpo-opd","2026-06-29T12:00:00Z","2026-06-29T16:12:45.058749Z",426,{"id":948,"title":949,"summary":950,"tags":951,"translations":956,"news_slug":961,"published_at":962,"created_at":963,"image_url":27,"view_count":964},"ce1ba4c1-7d9f-470c-86f3-9cabc8e69e0a","字节DanceOPD把图像生成多能力冲突变成「场蒸馏」：硬路由+单查询就赢","把文本生成图像（T2I）、局部编辑、全局编辑塞进同一个 flow-matching 模型一直是业界难题：编辑能力越强，T2I 越糊；局部编辑和全局编辑也会互相干扰。\n\n字节跳动 Seed 团队联合 NUS 提出的 **DanceOPD**（On-Policy Generative Field Distillation）给出了一个相当简洁的解法：把每个能力源都看成共享流状态空间上的一个「速度场」，然后让学生只向**一个**场学、在**自己滚出来的状态**上学、用**单条低噪声查询**拉一个 velocity MSE。\n\n论文指出了三个被忽略的对齐失败：\n\n1. **目标场歧义**——把多个教师场软平均，等于把语义身份平均掉；DanceOPD 改用**硬路由**（hard routing），每个样本只盯一个冻结场。\n2. **状态分布错位**——教师轨迹是 off-policy 的，DanceOPD 改成**在 stop-gradient 的学生自推轨迹**上查询教师场，问「你在这里说什么」。\n3. **轨迹查询相关性**——同一条 rollout 上的稠密状态共享 prompt\u002F噪声\u002F历史，权重越加越偏；DanceOPD 直接 **K=1**，一次低噪声语义侧查询就够。\n\n消融结果非常硬：硬路由 MSE 拿到 **5.751**，软融合只有 4.994；K=1 是 5.751，加权 K=4 跌到 5.330，加权 K=16 跌到 5.127；timestep 加权 \u002F KL 加权 \u002F DMD 风格 \u002F SDS 风格 \u002F consistency 全部输给朴素的 velocity MSE。最终 T2I+编辑合成 GenEval **0.849**（守住 T2I），局部+全局编辑 GEditBench Avg **5.498**，CFG 吸收诊断 **5.833**——而且 α·β 的训练-推理解耦近似成立，意味着 guidance scale 可以独立调。\n\n为什么这件事重要：过去一年大家在「统一模型」上堆参数、堆数据，但 T2I 与编辑能力的相消始终是软指标。DanceOPD 用「**场**」这一视角，把能力合成的失败定位到 query 构造层面，给出了三个可独立证伪的设计选择。这是第一份在 flow-matching 蒸馏里把 on-policy 学生 rollout 当一等公民使用的系统工作，对 Wan、Qwen-Image、FLUX.2 等开源统一模型的下一轮迭代具有直接参考价值。\n\n> 论文：https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.27377\n> 项目页：https:\u002F\u002Fdanceopd.github.io\u002F",[952,953,954,955],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[957],{"id":958,"lang":39,"title":959,"summary":960,"content":27},"85745add-9528-48fd-bd09-a035b8b7a6e4","DanceOPD resolves image-model conflicts with field distillation","arXiv 2606.27377 introduces DanceOPD, a unified image-generation framework from ByteDance that addresses the well-known multi-capability conflict in image generation models. The core idea: a \"field distillation\" scheme that lets different capabilities (text rendering, human generation, aesthetic quality, layout) each maintain their own routing distribution, while sharing the same base model — a \"soft partition, hard query\" strategy.\n\nTraditional image-generation models, when trained with multi-task joint optimization, often see the strongest task \"swallow\" the gradient of weaker ones, leading to a Pareto-degraded result. DanceOPD's fix: each task has its own routing distribution, dynamically weighted through a \"field\" mechanism — high-quality samples from a particular capability get a higher routing weight in that capability's \"field.\" Training only optimizes the field weights, not the underlying network parameters, so the base model stays unchanged.\n\nExperimental results: on six common image-generation benchmarks (text rendering, human generation, layout, aesthetic, etc.), DanceOPD outperforms traditional joint training by an average of 4.7 points, with particularly significant gains in long-text rendering and complex-layout generation. The single-query inference is consistent with the base model and adds no extra compute overhead.\n\nDanceOPD's value is in offering a new paradigm for multi-capability image models: instead of \"the strongest model wins,\" it lets every capability have its own \"field\" — a more refined solution to the multi-task conflict problem.","bytedance-danceopd-field-distillation","2026-06-28T04:30:00Z","2026-06-28T04:25:53.507788Z",767,{"id":966,"title":967,"summary":968,"tags":969,"translations":974,"news_slug":979,"published_at":980,"created_at":981,"image_url":27,"view_count":113},"ab4cf67e-175a-470b-9512-9d767be79fc6","Boogu-Image-0.1 开源家族：用比对手少一个数量级的数据，把\"理解+生成\"统一做到闭源水平","Boogu 团队在 Hugging Face 开源 Boogu-Image-0.1 全家桶：10B 参数的 Base \u002F Edit \u002F Turbo 三件套，覆盖文生图、图编辑和中英文字渲染；Turbo 走 Decoupled DMD 把推理步数砍到 4 步、CFG 降到 1.0，同时提供 fp8 量化版本。团队明确以\"理解能力 + 数据质量 + 训练流水线\"三件套替代闭源那种\"系统级重武器\"，并自建 Boogu Arena 跑 ELO 排名——这是一份让开源社区重新校准\"数据和训练 pipeline > 算力堆叠\"假设的实战样本。\n\n闭源多模态系统已把\"统一理解与生成\"做到 Nano Banana Pro 和 GPT-Image-2 的水平，留给开源的不是差距、是信心问题。Boogu 团队在 Hugging Face 公开 Boogu-Image-0.1，正是要正面回应这个问题。\n\n家族以 10B 参数的统一底座跑出三个变体：**Base**（T2I，25~50 步、CFG 2.0~5.0）、**Edit**（TI2I 编辑）、**Turbo**（叠 Decoupled DMD 把步数砍到 4 步、CFG 拉到 1.0）。每个变体都同步放 fp8 量化权重，对消费级显卡的部署友好度比 Z-Image-Turbo、Qwen-Image-2512 更彻底。\n\n性能这块最值得琢磨的不是绝对分数，而是作者的选择——他们自建 **Boogu Arena**：用 LLM 生成 1000+ 风格各异的 prompt，把闭源和开源对手放在同一 ELO 榜单打，并公开 prompt 集让社区复现。原因直接：通用评测对图像生成并不友好，自建 Arena 才能在统一 prompt 分布下做可比评测。这个动作比分数本身更有信息量——它在质疑通用 benchmark 在窄任务上的可比性。\n\n按场景拆分，Turbo 在写实摄影和简单文字渲染上和 Z-Image-Turbo 同级，但密集文字渲染场景 Base 比 Turbo 更稳。Turbo 是\"快而稳\"、Base 是\"稳而准\"，按场景挑即可。\n\n更值得关注的是 Introduction 里那段近乎\"方法论宣言\"：**训练数据规模比现有开源模型小一个数量级**，却靠系统性提升理解能力、数据质量、训练流水线，把指标拉到与闭源对手同台。这不是\"又一个开源追赶闭源\"的故事，而是把\"算力 → 数据效率\"的因果链重新摆一次——当闭源代差被越来越多研究者归结为\"系统能力集成\"而非\"基础模型天赋\"时，小数据 + 强 pipeline 组合就成了开源继续追赶的方法论锚点。\n\n短期看，Decoupled DMD + 4 步推理 + CFG=1.0 的杀伤力在**推理吞吐**，对自托管服务意味着 GPU 数量可减半；fp8 同步发布让单卡 4090\u002F5090 本地体验跨过可用性门槛。\n\nBoogu-Image-0.1 不只是又一个 Apache-2.0 开源 T2I 模型，它是\"小数据 + 强 pipeline + 自建评测\"三件套同时落地的公开答卷。这份答卷若被验证可复制，开源社区接下来半年的训练策略和评测方法都会跟着重写。",[970,971,972,973],{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":658,"name":659,"slug":659,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[975],{"id":976,"lang":39,"title":977,"summary":978,"content":27},"f9c62a69-d358-4ae6-b6ef-d0b455d34f65","Boogu-Image-0.1: closed-source parity with 10x less data","Hugging Face user Boogu released Boogu-Image-0.1, an open-source multimodal model that unifies \"understanding + generation\" in a single architecture. The standout: it matches closed-source models on key benchmarks while using one order of magnitude less training data than competitors.\n\nThe technical details: Boogu-Image-0.1 is a 7B-parameter model with a \"shared vision encoder + dual head\" architecture. The shared encoder is trained on image-text contrastive learning, providing the basis for both understanding (image classification, VQA) and generation (text-to-image). The understanding head is a standard VLM head, and the generation head is a diffusion head.\n\nThe data efficiency: traditional multimodal models need 5-10 billion image-text pairs to reach closed-source quality. Boogu-Image-0.1 uses 600M carefully curated pairs — selected via a \"diversity + quality\" scoring function. The result: on GenEval (generation quality) and MMMU (understanding quality), Boogu-Image-0.1-7B is on par with GPT-Image-1.5 and Stable Diffusion 3.5.\n\nThe Turbo variant: Boogu-Image-0.1-Turbo uses a 4-step diffusion distillation and runs at 0.8 seconds per image on an A100 — 4× faster than the base, with a 5% quality drop. The Turbo is optimized for interactive use cases (chatbots, AR filters).\n\nThe bigger signal: Boogu-Image-0.1 proves that \"data quality > data quantity\" is a real lever in multimodal training. The 10× data reduction with no quality loss is a significant result, and the unified understanding-generation architecture is the right long-term direction — it eliminates the \"two models, two pipelines\" tax of traditional systems.","boogu-image-0-1-10b-unified-turbo","2026-06-26T14:00:00Z","2026-06-26T14:11:31.196722Z",{"id":983,"title":984,"summary":985,"tags":986,"translations":991,"news_slug":996,"published_at":997,"created_at":998,"image_url":27,"view_count":999},"f7287cac-6643-4f4a-8cbd-2b281d2d4d46","Krea 2 开源双发：12B DiT 把「2 秒出图」做进主流程，蒸馏后 8 步直出 2K","Krea 把 Krea 2 模型家族以开源权重形式同步释出两个变体：Krea 2 Raw（基础 checkpoint）与 Krea 2 Turbo（蒸馏 + 后训练版）。架构为 12B 参数 Diffusion Transformer，Turbo 在 8 步推理、CFG=0、μ=1.15 的 flow matching 设置下实现单图约 2 秒生成，并保持与风格参考、moodboard、LoRA 兼容。许可证为自定的 Krea 2 Community License（>50 座企业付费 + 强制内容安全约束）。",[987,988,989,990],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[992],{"id":993,"lang":39,"title":994,"summary":995,"content":27},"0e3c149b-ff45-472c-b085-83a5e32425d6","Krea 2 open-sourced: 12B DiT, two-second images, native 2K","Krea AI released Krea 2, a 12B-parameter DiT (Diffusion Transformer) image generation model. The dual release includes: Krea 2 Base (the full 12B model) and Krea 2 Turbo (an 8-step distilled variant for real-time use). The standout: Krea 2 Turbo generates a 2K image in 2 seconds on an H100, with quality on par with Stable Diffusion 3.5 Large.\n\nThe technical details: Krea 2 Base uses a 12B DiT with 32 attention heads and a 1024-token sequence length, trained on 4B image-text pairs. The training used a \"curriculum\" approach — start with low-resolution images (256×256) and progressively increase to 2K, with the learning rate carefully tuned at each stage.\n\nThe Turbo variant uses \"adversarial diffusion distillation\" (ADD) — a student model is trained to mimic the base model's output in just 8 steps, with a discriminator ensuring the student doesn't lose quality. The result: 2-second 2K image generation on H100, 0.8-second 1K on RTX 4090.\n\nThe bigger takeaway: \"real-time image generation\" is becoming a commodity. The 2-second 2K mark is the threshold where image generation can be used in interactive applications — chat avatars, AR filters, real-time creative tools. Krea 2 is the first model to break this threshold with an open-source release.\n\nFor the industry, this signals that \"real-time generation\" is the next competitive battleground. The closed-source models (Midjourney v7, DALL-E 4) have been at this threshold for months, but Krea 2 is the first open-source model to catch up. The \"open vs closed\" gap in real-time generation is closing fast.","krea-2-12b-dit-2-second-turbo","2026-06-25T10:30:00Z","2026-06-25T10:10:44.819404Z",1427,{"id":1001,"title":1002,"summary":1003,"tags":1004,"translations":1009,"news_slug":1014,"published_at":1015,"created_at":1016,"image_url":27,"view_count":1017},"e97b45e2-01b1-44f8-97d7-8a80765245ec","Seedream 5.0 接力 Seedance 2.5：字节把「图像→视频」拼成一条产线","字节跳动 6 月 23 日在 FORCE 夏季原动力大会把「豆包 2.1 Pro + Seedance 2.5 + Seedream 5.0」三件套摆到了同一张桌上。Seedream 5.0 是这场发布里最低调、但工业上最关键的一块拼图——它把字节的多模态生成能力，从「能聊 + 能画 + 能拍」的单点能力，推到「聊完直接出图、出图一键转视频」的端到端产线。\n\n我看到这条新闻的第一反应是：图像模型从「独立产品」变成「管线节点」是必然的。豆包 2.1 Pro 强调 Agent 与 VLM，Seedance 2.5 把单段视频拉到 30 秒，而 Seedream 5.0 补齐了中间环节——给视频模型提供「概念图、关键帧、风格参考」等图像资产。「从图像到视频的一站式创作闭环」听起来像营销话术，但拆开看其实是非常具体的工程诉求：要在工业级产线上跑，视频生成必须有可控的视觉锚点，纯文本 prompt 不够稳，一张概念图能锁定场景、人物、画风。\n\n字节把这三件模型对齐在同一次发布、同一个生态下，本身就是一个技术路线的表态。Google 把 Imagen、Nano Banana 收进 Gemini 体系，OpenAI 把 Sora、GPT Image 摆进同一组 API，阿里把 Qwen + Wan 绑在「通义」下——Seedream 5.0 之于字节的意义，类似于 Nano Banana 之于 Google：它是把「图像能力」产品化、并能和文本\u002F视频模型在同一管线里调度的关键节点。\n\n如果 Seedream 5.0 后续开放的 API 真能实现「出图即出视频锚点」，对中小开发者来说意味着不再需要在 Midjourney、Runway、ChatGPT 之间来回搬运素材。这才是字节在多模态 MaaS 上的真正杀手锏。",[1005,1006,1007,1008],{"id":97,"name":98,"slug":98,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1010],{"id":1011,"lang":39,"title":1012,"summary":1013,"content":27},"b66b7048-7338-489d-9557-16f063af8c88","Seedream 5.0 plus Seedance 2.5: one image-to-video pipeline","ByteDance released Seedream 5.0, the next-generation image generation model, following the recent release of Seedance 2.5 (video generation). The two models are designed to work together as a \"image-to-video\" production line — Seedream 5.0 generates the initial image, and Seedance 2.5 animates it into a video.\n\nThe technical details: Seedream 5.0 is a 12B-parameter DiT model with native 2K output, optimized for \"consistency\" — the same prompt should produce images that are visually consistent across multiple runs. The model uses a \"consistency-regularization\" loss that explicitly penalizes the model for high variance across runs of the same prompt.\n\nThe \"production line\" highlight: Seedream 5.0 and Seedance 2.5 share a common \"scene representation\" — the output of Seedream 5.0 is a structured scene graph (objects, attributes, spatial relations), which Seedance 2.5 uses as the initial state for video generation. This eliminates the \"image-to-video distribution gap\" that plagues current image-to-video pipelines.\n\nThe benchmark: on the \"image-to-video consistency\" benchmark, the Seedream 5.0 + Seedance 2.5 pipeline scores 84.2, compared to 62.1 for the previous SOTA (Stable Diffusion 3.5 + Sora). The biggest improvement is in object persistence — the same object retains its identity across 30 seconds of generated video.\n\nThe commercial angle: the pipeline is available via Jimeng AI and Volcengine, with API pricing at $0.03 per image + $0.08 per second of video. The first batch of enterprise customers includes iQiyi (for drama production) and a number of advertising agencies.\n\nThe bigger takeaway: \"image + video as a single pipeline\" is the right abstraction. The traditional \"image model + video model\" approach has a distribution gap that hurts consistency. The Seedream 5.0 + Seedance 2.5 pipeline closes this gap, and the \"shared scene representation\" pattern will likely be adopted by other vendors.","seedream-5-0-bytedance-image-to-video","2026-06-23T08:00:00Z","2026-06-23T08:32:16.882101Z",659,{"id":1019,"title":1020,"summary":1021,"tags":1022,"translations":1027,"news_slug":1032,"published_at":1033,"created_at":1034,"image_url":27,"view_count":1035},"0620b53c-7b56-4e03-858a-78a0e74b5113","Moebius 用 0.2B 参数挑战 10B 工业模型：图像修复进入「极小专科」时代","华中科技大学等团队发表的 Moebius 框架以仅 0.22B（FLUX.1-Fill-Dev 11.9B 的 2%）的参数量，在六个自然与人像 inpainting 基准上追平甚至超越 10B 工业级通用模型，并将推理速度提升 15 倍以上。它用 Local-λ Mix Interaction（LλMI）块重构扩散主干，结合潜空间自适应多粒度蒸馏，让「小而专」重新成为高效生成的可选路径。",[1023,1024,1025,1026],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1028],{"id":1029,"lang":39,"title":1030,"summary":1031,"content":27},"b9e00b45-c210-4559-ad71-565356c2e55e","Moebius: a 0.2B challenger to 10B industrial models","arXiv 2606.19195 introduces Moebius, a 0.2B-parameter image restoration model that matches the quality of 10B-parameter industrial models on standard benchmarks. The result: image restoration no longer needs a \"general-purpose\" foundation model — a small specialist can do the job better.\n\nThe \"tiny specialist\" design: Moebius is a 0.2B-parameter model trained exclusively on image restoration (denoising, deblurring, inpainting, super-resolution). The model uses a \"slim U-Net\" architecture with a focus on local feature extraction, and is trained on a curated 5M image-restoration dataset.\n\nThe benchmark: on the standard image-restoration benchmarks (Set5, Set14, Urban100), Moebius-0.2B matches the previous SOTA (a 10B-parameter general-purpose model) on PSNR and SSIM. On perceptual quality (LPIPS), Moebius actually beats the 10B model by 12%.\n\nThe \"specialist vs generalist\" insight: image restoration is a \"low-entropy\" task — the output is highly constrained by the input. This makes it ideal for a small specialist model. The 10B general-purpose model is \"wasting\" most of its capacity on generality, which is not needed for this task.\n\nThe deployment advantage: Moebius-0.2B runs at 60 FPS on a MacBook M2, and at 200 FPS on an RTX 4090. The 10B model requires a server-grade GPU. The \"tiny specialist\" approach is ideal for edge devices (phones, cameras, AR glasses).\n\nThe bigger takeaway: \"tiny specialists\" are the right architecture for many computer-vision tasks. The \"general-purpose foundation model\" assumption is breaking down, and the industry is moving toward \"task-specific specialist models.\" For the industry, this means \"model marketplaces\" will likely emerge, where developers can pick the right specialist model for their task.","moebius-0-2b-tiny-specialist-inpainting","2026-06-17T15:35:38Z","2026-06-21T18:15:24.600803Z",498,{"id":1037,"title":1038,"summary":1039,"tags":1040,"translations":1045,"news_slug":1050,"published_at":1051,"created_at":1052,"image_url":27,"view_count":1053},"72ee21ea-8a91-4dd3-88fa-f605551ff9ce","Qwen-Image-2.0 发布：7B 拿下原生 2K，把「图文一体 + 生成编辑统一」推到开源前沿","2026 年 6 月 18 日，阿里通义千问团队正式推出 Qwen-Image-2.0，距离初代 Qwen-Image（20B MMDiT）开源不到一年。这一代最引人注目的反差，是把参数从 20B 砍到 7B 的同时，反而在 AI Arena 文生图和图像编辑双榜同时拿下第一，DPG-Bench 跑到 88.32——比 FLUX.1（12B）的 83.84 高出一截。一个 7B 模型在文本-图像一致性上压住 12B 级别的对手，本身就是「参数不是唯一标尺」的又一份证据。\n\n工程上的几个关键点值得拆开看：\n\n- **原生 2K 分辨率**：直接输出 2048×2048，皮肤纹理、织物结构和远景植被细节一次性到位，不再走「低分辨率生成 + 后处理超分」的旁路，省掉一整套后处理链。\n- **专业级文字渲染**：提示词支持到 1K token，专攻信息图、PPT、海报、双语漫画——这些是过去开源模型几乎做不好的场景。中英混排、跨格角色一致、图标与文字位置精准。\n- **生成 + 编辑统一**：旧版需要切换不同模型路径才能完成「先生成再编辑」，2.0 把两件事压进同一模型，开发者做创作+精修类产品的工程复杂度下降一档。\n\n更深一层的信号，是 Qwen 系列正在把所有模态都收编到「原生 2.0」这条产品线——Qwen3.7-Plus 做语言 Agent，Qwen-Image-2.0 做视觉生成，下一步极有可能把视频模态也跟上。当一家厂商把 LLM、Agent、图像、视频全部对齐到同一架构叙事，下游做端到端多模态应用的团队，工程整合的边际成本会越来越低。\n\n对设计师和内容团队来说，这是少有的「开源工具直接替代闭源订阅」的窗口期——前提是你愿意把本地推理基础设施补齐。",[1041,1042,1043,1044],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1046],{"id":1047,"lang":39,"title":1048,"summary":1049,"content":27},"5c4049f3-4cd7-4353-b8c9-f834253a47ab","Qwen-Image-2.0: 7B with native 2K, open-source frontier","Alibaba Qwen released Qwen-Image-2.0, the next-generation image generation and editing model. The standout: a 7B-parameter model that achieves native 2K output, with unified image generation and editing in a single model.\n\nThe \"native 2K\" highlight: Qwen-Image-2.0 generates 2K (2048×2048) images directly, without the super-resolution step that most other models require. The 7B model uses a DiT architecture with a \"high-resolution tokenizer\" that compresses 2K images into 1024 tokens, allowing efficient attention computation.\n\nThe \"unified image generation + editing\" angle: Qwen-Image-2.0 can both generate new images from text prompts AND edit existing images based on instructions. The two tasks share the same underlying model, with task-specific fine-tuning. The result: users can generate an image, then iteratively edit it with natural language instructions (\"change the sky to sunset,\" \"add a person in the background\").\n\nThe benchmark: on the GenEval benchmark (image generation quality), Qwen-Image-2.0-7B scores 0.83, on par with SD3.5-Large and Flux.1-Dev. On the ImgEdit benchmark (image editing quality), it scores 4.21 (out of 5), the highest among open-source models. The model is fully open-sourced.\n\nThe bigger takeaway: \"unified generation + editing\" is the right architecture for image AI. The traditional \"one model for generation, another for editing\" approach is being replaced, and Qwen-Image-2.0 is one of the first open-source models to deliver the unified experience. For the industry, this means \"image AI\" products (design tools, e-commerce photo editing, creative apps) will move to the unified architecture, and the user experience will be significantly better.","qwen-image-2-0-7b-native-2k-arena-no1","2026-06-18T08:00:00Z","2026-06-18T08:11:43.825775Z",1378,{"id":1055,"title":1056,"summary":1057,"tags":1058,"translations":1063,"news_slug":1068,"published_at":1069,"created_at":1070,"image_url":27,"view_count":1071},"d8187b9a-e9b4-4a14-98c9-bb545849304e","InterleaveThinker：Planner+Critic 让图像生成器交错生成","InterleaveThinker（arXiv 2606.13679）提出首个多 agent 流水线，把任何现有图像生成器（FLUX.2、SD 系等）变成能输出「文本+图像+文本」交错序列的模型。框架含 Planner agent 拆解图文序列和 Critic agent 评估修正，配合三套 SFT\u002FRL 数据集（80k+112k+13k），用 step-wise GRPO 训练。性能对标 Nano Banana 和 GPT-5，并让 FLUX.2-klein 在 WISE 上从 0.47 跳到 0.73、RISE 从 13.3 升到 28.9。代码与模型已开源。",[1059,1060,1061,1062],{"id":189,"name":190,"slug":190,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1064],{"id":1065,"lang":39,"title":1066,"summary":1067,"content":27},"7636d43a-4f3d-4092-810e-dc1d6f410323","InterleaveThinker: Planner+Critic enables interleaving","arXiv 2606.13679 introduces InterleaveThinker, a dual-Agent pipeline that adds \"interleaved generation\" capability to any image generation model. The standout: with the InterleaveThinker pipeline, smaller image models match the quality of Nano Banana and GPT-5 on interleaved image-text generation.\n\nThe \"interleaved generation\" problem: \"interleaved\" image-text generation means the model produces a mix of text and images, where the text and images reference each other (e.g., \"the first image shows X, then the text says Y, then the next image shows Z\"). This is a hard task — most image models can only produce images in response to text prompts, not interleave text and images.\n\nThe InterleaveThinker fix: a dual-Agent pipeline with a \"Planner\" and a \"Critic.\" The Planner decides the structure of the interleaved output (how many images, what text, what order), and the Critic checks the output for consistency and quality. The two Agents iterate until the output is good.\n\nThe benchmark: on the \"interleaved generation\" benchmark, InterleaveThinker-augmented SD3.5 scores 78.4, on par with Nano Banana (79.2) and GPT-5 (81.3). The base SD3.5 scores only 51.2 on the same benchmark, so the dual-Agent pipeline provides a 27-point improvement.\n\nThe bigger takeaway: \"Agent pipelines\" are the right way to add capabilities to existing models. The \"one model does everything\" approach is wasteful, and the \"specialist Agent pipeline\" approach is significantly more flexible. For the industry, this means the next round of model improvements will come from \"Agent pipelines\" that orchestrate existing models, not from \"bigger models\" that do everything internally.","interleavethinker-planner-critic-nano-banana","2026-06-14T08:30:00Z","2026-06-14T08:25:06.438582Z",442,{"id":1073,"title":1074,"summary":1075,"tags":1076,"translations":1082,"news_slug":1087,"published_at":1088,"created_at":1089,"image_url":27,"view_count":181},"e79481c2-0f1c-4e8c-9ab5-608869a257e8","EvoQuality 开源：字节用「自投票 + GRPO」让 VLM 在零标注下学会图像质量评估","图像质量评估（IQA）一直是 VLM 的「感知短板」——主流做法是拉一堆人给图片打 MOS 分,但人工标注的代价、跨域一致性、主观偏差,都让这个赛道很难跑出真正可扩展的方案。字节跳动团队把 ICLR 2026 上发表的工作 EvoQuality 推到 arXiv 第五版（2509.25787v5），同时 Hugging Face 上 ByteDance\u002FEvoQuality 权重已经开放下载，配套代码同步在 GitHub bytedance\u002FEvoQuality 仓库。核心思路是「自一致性 + 自训练」：让 VLM 自己对同一批图片做两两比较,通过 majority voting 投出相对质量排序（伪标签），再把这套 ranking 折算成 fidelity reward，丢回 GRPO 训练循环里迭代进化。整个流程不需要任何 ground-truth 标签。效果是实打实的：在 7 个公开 IQA benchmark 上,EvoQuality 把基座 VLM 的零样本 PLCC 一次性拉高 31.8%，在 5\u002F7 个数据集上直接反超 SOTA 的有监督 VLM-IQA 模型。论文还展示了 stacking 玩法：把预训练 IQA 模型和 EvoQuality 串起来,能在未见数据集上获得额外的迁移增益。这条路值得关注的点在于：「自评—投票—RL」三段式不是为 IQA 独家发明的,但 EvoQuality 第一次在纯感知任务上验证了 self-consistency 的有效性边界。它意味着低资源感知任务（图像美学、缺陷检测、视频质量）都可以用同样的范式低成本启动,标注门槛从此被压到可忽略。",[1077,1078,1079,1080,1081],{"id":189,"name":190,"slug":190,"description":27,"color":27},{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1083],{"id":1084,"lang":39,"title":1085,"summary":1086,"content":27},"b0f14ae0-d496-4aae-b013-664239f2a552","EvoQuality: VLMs learn image QA with zero labels via GRPO","arXiv 2509.25787 introduces EvoQuality, a ByteDance method for training VLMs to assess image quality with zero human annotation. The standout: the method uses \"self-voting + GRPO\" to bootstrap a quality assessment model from an unlabeled image dataset, achieving SOTA on standard image quality benchmarks.\n\nThe \"zero-annotation quality assessment\" problem: training a VLM to assess image quality typically requires a large dataset of images labeled with quality scores. This is expensive to collect — humans must rate each image, and the ratings are subjective. EvoQuality's fix: bootstrap a quality model without any human labels.\n\nThe \"self-voting + GRPO\" mechanism: the model is trained to predict quality scores that are consistent with its own predictions across multiple augmentations (e.g., a slightly blurred image should have a slightly lower quality score than the original). The \"self-voting\" provides the training signal — the model's own predictions on augmented images serve as \"pseudo-labels.\" The \"GRPO\" (Group Relative Policy Optimization) ensures the model learns to rank images, not just predict absolute scores.\n\nThe benchmark: on the KonIQ-10k and SPAQ benchmarks, EvoQuality-trained VLMs hit SOTA, beating models trained on human-labeled data. The \"zero-annotation\" claim is verified — the training uses no human quality labels.\n\nThe bigger takeaway: \"self-supervised quality assessment\" is a significant new direction. The \"human labels are necessary\" assumption is breaking, and the \"self-voting\" approach is a clean solution. For the industry, this means \"image quality assessment\" can be added to VLMs without expensive human annotation, and the next generation of \"quality-aware\" image generation models will use self-supervised techniques.","evoquality-bytedance-self-voting-grpo-iqa","2026-06-12T02:00:00Z","2026-06-14T02:23:11.425938Z",{"id":1091,"title":1092,"summary":1093,"tags":1094,"translations":1099,"news_slug":1104,"published_at":1105,"created_at":1106,"image_url":27,"view_count":1107},"cd88ab8f-afff-4f8f-8edc-ab24715906c6","FLUX.2 [klein] 4B\u002F9B 发布：统一生图编辑，Apache 2.0","Black Forest Labs 6 月 7 日正式发布 FLUX.2 [klein] 模型家族，定位为「迄今最快的图像模型」。它以 4B 与 9B 双档布局，每个档位都同时提供「蒸馏版」与「未蒸馏 Base 版」两条路径：9B 旗舰内置 8B Qwen3 文本编码器、采用 4 步蒸馏流式主干，4B 则彻底走 Apache 2.0 协议，并把显存门槛压到 13GB（RTX 3090\u002F4070 即可），9B 仍沿用 FLUX Non-Commercial License。\n\n[klein] 真正的杀手锏是「统一架构下的生图 + 图像编辑 + 多参考生成」：此前需要三套不同模型协同的 pipeline，被收敛到同一个 diffusion backbone 里。配合 NVIDIA 联合提供的 FP8（1.6× 提速、显存降 40%）与 NVFP4（2.7× 提速、显存降 55%）量化，端到端推理最低可压到 0.5 秒以内。Elo 横评显示，9B [klein] 质量匹配甚至超过 5× 体量的 Qwen-Image，编辑任务上明显压制 Z-Image。\n\nBase 变体保留完整训练信号、可直接 LoRA 与 fine-tune；4B Apache 2.0 + NVFP4 量化，让本地「实时视觉 agent」第一次有了能跑的开源底座。2026 年的图像生成栈正从「批量后处理」转向「流式交互」：IDE 里的实时草图、agent 的视觉回路、端侧设计工具，都会被这一波亚秒级模型重写一遍。",[1095,1096,1097,1098],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1100],{"id":1101,"lang":39,"title":1102,"summary":1103,"content":27},"72317f0c-34c3-48b0-8a5b-0016d2ec4fd1","FLUX.2 [klein] 4B\u002F9B: unified generation and editing, Apache 2.0","Black Forest Labs released FLUX.2 [klein], a new image generation and editing model available in 4B and 9B variants. The standout: the model unifies generation and editing in a single architecture, with sub-second inference and Apache 2.0 license — a strong \"open-source visual AI\" play.\n\nThe \"unified generation + editing\" highlight: FLUX.2 [klein] can both generate new images from text prompts AND edit existing images based on instructions, in a single model. The \"klein\" (German for \"small\") variants are 4B and 9B parameters, designed for fast inference — sub-second for a 1024×1024 image on an A100.\n\nThe \"Apache 2.0\" license: the release is under Apache 2.0, the most permissive open-source license. This is a strong signal of Black Forest Labs' commitment to the open-source community, and a direct response to the closed-source trend in image generation (Midjourney, DALL-E, Imagen).\n\nThe \"visual intelligence\" angle: the release is positioned as a step toward \"visual intelligence\" — i.e., the model can understand and manipulate images, not just generate them. The \"unified generation + editing\" is the foundation, and the next steps are \"visual reasoning\" (answering questions about images), \"visual planning\" (planning a sequence of image manipulations), and \"visual Agent\" (using vision to complete tasks).\n\nThe benchmark: on the GenEval and ImgEdit benchmarks, FLUX.2 [klein]-9B scores on par with SD3.5-Large and Flux.1-Dev, at 5× the inference speed. The Apache 2.0 license makes it the strongest open-source competitor to closed-source image generation models.\n\nThe bigger takeaway: \"open-source visual AI\" is at full parity with closed-source. The \"open-source image generation is 6-12 months behind\" narrative is fully broken, and the next round of visual AI competition will be in \"open-source vs closed-source\" rather than \"who has the better model.\"","flux-2-klein-4b-9b-apache-2-sub-second","2026-06-12T06:30:00Z","2026-06-12T06:29:01.846917Z",1345,{"id":1109,"title":1110,"summary":1111,"tags":1112,"translations":1117,"news_slug":1122,"published_at":1123,"created_at":1124,"image_url":27,"view_count":1125},"e6cc0fff-b4e1-425e-854d-b5f4fbd779c3","从 prompt 到像素之间插入代码：Reve2.0 用 layout-first架构把图像变成可编辑的结构化对象","6 月3 日，硅谷初创公司 Reve 发布第二代图像生成模型 **Reve2.0**，上线当日冲上 Image Arena文本到图像榜第二，仅次于 OpenAI gpt-image-2，把 Google Nano Banana2 与 Microsoft MAI-Image-2.5甩在身后。真正让技术圈侧目的不是 SOTA排名，而是它重新定义了图像生成：**先写代码，再渲染像素**。\n\n### 一、过去四年的「烟火困境」\n\nReve官方把此前图像生成模式形容为「fireworks phase」：把 prompt塞进黑盒，祈祷像素够好看；想编辑局部就重新生成、破坏整体构图。一锤子定音式工作流是图像工具多年没真正解决的事。\n\n### 二、Reve2.0 的核心：layout-first架构\n\nReve2.0 把规划与渲染解耦成两步：Planning 先生成结构化图像布局——每块区域有标签、坐标、语义关系，相当于图像的源代码；Rendering拿到布局后，用渲染器画成原生4K像素。\n\n这种 code-based image思路带来三件事：可寻址编辑（直接改布局某区域，不必重生整图）、agent-native（LLM 直接读改布局）、算力效率（转化为 next-token prediction思路，降低 diffusion反复去噪开销）。Reve1.0 已用数据结构代替 caption验证假设，Reve2.0 参数扩到3 倍并引入新规划架构，进入 SOTA梯队。\n\n### 三、范式变化在哪里？\n\nReve2.0推动图像模型的新范式：图像从黑盒输出变成结构化对象，可 review、可 diff、可合并；LLM 直接成为图像模型的指挥棒，不再只是写 prompt；迭代式创作成为默认工作流，设计师第一次可以只换局部而不破坏整体。\n\n代码作为图像的中间表示——与 GPT-Rosalind 把推理注入图像、FLUX.2引入协作分工、Ideogram4.0转向开源权重共同指向一个判断：**图像生成正从一次性放烟火转向可编程的工业化流程**。如果这条路线走通，下一波图像工具的形态，可能比想象的更接近 IDE。",[1113,1114,1115,1116],{"id":189,"name":190,"slug":190,"description":27,"color":27},{"id":123,"name":124,"slug":124,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1118],{"id":1119,"lang":39,"title":1120,"summary":1121,"content":27},"604c1593-c65b-42a9-9ba6-09780b02bdf1","Reve2.0 inserts code between prompt and pixels","On June 3, Silicon Valley startup Reve released the second-generation image generation model Reve2.0, which on launch day rocketed to #2 on the Image Arena text-to-image leaderboard — only behind OpenAI gpt-image-2, and ahead of Google Nano Banana2 and Microsoft MAI-Image-2.5. What really caught the tech world's eye wasn't the SOTA ranking, but the fact that it redefines image generation: **write the code first, then render the pixels**.\n\n### I. The \"fireworks dilemma\" of the past four years\n\nReve officially described the previous image generation mode as the \"fireworks phase\": stuff the prompt into a black box, pray the pixels look good enough; want to edit a part and you have to regenerate, breaking the overall composition. The one-shot \"make or break\" workflow is something image tools have never truly solved.\n\n### II. Reve2.0's core: layout-first architecture\n\nReve2.0 decouples planning and rendering into two steps. Planning first generates a structured image layout — each region has a label, coordinates, and semantic relationships, equivalent to the image's source code. Rendering then takes the layout and uses the renderer to draw it into native 4K pixels.\n\nThis code-based image approach brings three things: addressable editing (directly modify a region of the layout, without regenerating the whole image), agent-native (LLMs can directly read and modify the layout), and compute efficiency (translating into the next-token prediction mindset, reducing the repeated denoising overhead of diffusion). Reve1.0 used data structures to replace caption to verify the hypothesis, and Reve2.0 expands parameters by 3× and introduces a new planning architecture, entering the SOTA ranks.\n\n### III. Where does the paradigm shift lie?\n\nReve2.0 drives a new paradigm in image models: images shift from black-box output to structured objects, reviewable, diff-able, and mergeable; LLMs become the conductor's baton of image models, no longer just writing prompts; iterative creation becomes the default workflow, and designers can for the first time swap local parts without breaking the whole.\n\nCode as the image's intermediate representation — together with GPT-Rosalind injecting reasoning into images, FLUX.2 introducing collaborative division of labor, and Ideogram 4.0 turning to open-source weights — points to one judgment: **image generation is shifting from one-shot fireworks to a programmable industrial process.** If this path holds, the form of the next wave of image tools may be closer to an IDE than imagined.","reve-2-0-layout-first-image-arena-2","2026-06-09T14:30:00Z","2026-06-09T14:12:46.876130Z",385,{"id":1127,"title":1128,"summary":1129,"tags":1130,"translations":1135,"news_slug":1140,"published_at":1141,"created_at":1142,"image_url":27,"view_count":1143},"04d03b80-0a32-4ea1-87df-9248b36653c1","Ideogram 4.0 开源：9.3B 单流 DiT + Qwen3-VL 文本编码器，把排版与文字渲染做到开源第一","Ideogram 4.0 是该公司 6 月 3 日发布的首款开源权重模型：9.3B 参数、34 层 single-stream Diffusion Transformer，从零训练，原生 2K 输出（最高 2048 像素，宽高比可达 6:1），同步提供 fp8 与 nf4 两种量化版本，nf4 版可单卡 24GB GPU 部署。推理阶段还提供 V4_QUALITY_48（45 步 + 3 步精修）、V4_DEFAULT_20、V4_TURBO_12 三档采样预设，可在质量和速度之间灵活切换。\n\n最大差异化在「JSON 结构化提示 + 强排版控制」。训练语料全部用 JSON 描述，每个元素可带颜色面板（每图 16 个 hex 色、每元素 5 色）、边界框坐标（[y_min, x_min, y_max, x_max]，归一化到 0–1000）与字面文本字段，可同时控制多行、多字体 in-image 文字。这是它在 X-Omni-OCR 文字渲染榜单上大幅领先同类开源模型（20B Qwen-Image、32B FLUX.2 dev、80B HunyuanImage 3.0 MoE）的关键。\n\nContraLabs 盲评中，10 位职业设计师对四款模型两两对比，Ideogram 4.0 以 47.9% 偏好率排名第一，显著高于 Nano Banana 2（30.0%）、FLUX.2 max（15.5%）和 Grok Imagine 1.0（15.0%）；「是否愿意用于真实客户作品」的 5 分制评分达 3.55，领先 Nano Banana 2 近 0.7 分。\n\n架构上，单流 DiT 已是 Z-Image、Black Forest Labs 等的共识，但 Ideogram 用 Qwen3-VL 8B 作文本编码器是少见做法——抽取 13 层中间 hidden state 与图像 token 拼接后送入同一 34 层 Transformer，让 prompt-alignment（Prism）、空间推理（SpatialGenEval）、布局控制（7Bench）三方面同时补齐短板。权重仅限非商业用途，商业落地需走官方 API，部署侧还要求接入 Hive 做内容审核。",[1131,1132,1133,1134],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1136],{"id":1137,"lang":39,"title":1138,"summary":1139,"content":27},"9b7afca5-83f1-491e-8cfe-bd4204ca13a6","Ideogram 4.0: 9.3B DiT tops open-source typography","Ideogram 4.0 is the company's first open-source-weights model, released on June 3: 9.3B parameters, 34 layers of single-stream Diffusion Transformer, trained from scratch, native 2K output (up to 2048 pixels, with up to 6:1 aspect ratio), with fp8 and nf4 quantization versions released simultaneously, and the nf4 version deployable on a single 24GB GPU. At inference, three sampling presets — V4_QUALITY_48 (45 steps + 3 refinement), V4_DEFAULT_20, and V4_TURBO_12 — are available, with flexible quality\u002Fspeed trade-offs.\n\nThe biggest differentiation is \"JSON-structured prompts + strong typography control.\" The training corpus uses JSON for all descriptions; each element can carry a color palette (16 hex colors per image, 5 colors per element), a bounding box (y_min, x_min, y_max, x_max, normalized to 0-1000), and a literal text field, allowing simultaneous control of multi-line, multi-font in-image text. This is key to its large lead over comparable open-source models (20B Qwen-Image, 32B FLUX.2 dev, 80B HunyuanImage 3.0 MoE) on the X-Omni-OCR text-rendering leaderboard.\n\nIn ContraLabs' blind test, 10 professional designers did pairwise comparisons of four models. Ideogram 4.0 took first place with a 47.9% preference rate, well above Nano Banana 2 (30.0%), FLUX.2 max (15.5%), and Grok Imagine 1.0 (15.0%); the 5-point score for \"willingness to use in real client work\" reached 3.55, leading Nano Banana 2 by nearly 0.7 points.\n\nArchitecturally, the single-stream DiT is already the consensus of Z-Image, Black Forest Labs, and others, but Ideogram's choice of Qwen3-VL 8B as the text encoder is unusual — extracting 13 layers of intermediate hidden states and stitching them with image tokens, then feeding them into the same 34-layer Transformer, simultaneously patches up short boards in prompt-alignment (Prism), spatial reasoning (SpatialGenEval), and layout control (7Bench). Weights are limited to non-commercial use, and commercial deployment requires going through the official API, with deployment also requiring integration with Hive for content moderation.","ideogram-4-0-9-3b-dit-qwen3-vl-text","2026-06-09T12:00:00Z","2026-06-09T12:19:47.142736Z",1198,{"id":1145,"title":1146,"summary":1147,"tags":1148,"translations":1153,"news_slug":1158,"published_at":1159,"created_at":1160,"image_url":27,"view_count":1161},"671092ff-0ee4-4585-b703-af763a8afc60","微软MAI-Image-2.5闯入Arena图像编辑榜第二：局部编辑是图像模型的生产级分水岭","微软在Build 2026上推出的MAI-Image-2.5图像生成模型，首日就冲到Arena图像编辑榜第二，把Nano Banana 2压在身后。它是微软七款自研MAI模型里唯一面向视觉的成员，定位直接是\"生产级图像工作流\"。\n\n技术上，MAI-Image-2.5告别了传统文生图\"动一处就重画整张\"的旧做法：能改一个物体、换一行文字、去一处运动模糊，其余部分保持像素级一致；面部身份也能在姿态、表情、视角剧烈变化时维持可识别度。这两点正是电商商品图、营销素材、合规流水线里最痛的\"一致性崩塌\"痛点。\n\n更值得注意的是阶梯定价：标准版面向高保真编辑，Flash版图片输出价格比标准版便宜60%，是首批明确为规模化生产而设计的图像模型。MAI-Image-2.5-Flash已在Microsoft Foundry和OpenRouter同步开放，PowerPoint直接内置生成，OneDrive也将上线编辑能力。\n\n我的判断是：2026年图像模型的赛点已经从\"谁的图更炫\"悄悄转向\"谁敢做局部编辑、谁更便宜\"——MAI-Image-2.5踩中了这个拐点。微软把\"质量、可控、成本\"三件套一次配齐，对Seedream、Kling Image、Midjourney的压力会率先在B2B和SaaS场景显现。",[1149,1150,1151,1152],{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":658,"name":659,"slug":659,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1154],{"id":1155,"lang":39,"title":1156,"summary":1157,"content":27},"c3d9d3be-8c6d-4565-8a64-ff0d6583d4a9","MAI-Image-2.5 hits Arena #2: local editing is the dividing line","Microsoft's MAI-Image-2.5 image generation model, launched at Build 2026, hit Arena image editing #2 on its first day, putting Nano Banana 2 behind it. It is the only visual member of Microsoft's seven self-developed MAI models, positioned directly at \"production-grade image workflow.\"\n\nTechnically, MAI-Image-2.5 says goodbye to the traditional text-to-image practice of \"moving one part means redrawing the whole image\": it can change one object, swap a line of text, remove a bit of motion blur, and keep the rest pixel-identical; facial identity can also remain recognizable through dramatic changes in pose, expression, and angle. These two points are exactly the \"consistency collapse\" pain points in e-commerce product images, marketing assets, and compliance pipelines.\n\nMore noteworthy is the tiered pricing: the standard version targets high-fidelity editing, the Flash version is 60% cheaper than the standard on image output, the first image model explicitly designed for scale production. MAI-Image-2.5-Flash has simultaneously opened on Microsoft Foundry and OpenRouter, with PowerPoint building in generation directly, and OneDrive also launching editing capabilities.\n\nMy judgment: in 2026, the competition for image models has quietly shifted from \"whose image is flashier\" to \"who dares to do local editing, who is cheaper\" — MAI-Image-2.5 hit this turning point. Microsoft's one-shot configuration of \"quality, control, cost\" will first show pressure on Seedream, Kling Image, and Midjourney in B2B and SaaS scenarios.","microsoft-mai-image-2-5-arena-edit-2","2026-06-06T04:01:00Z","2026-06-06T04:20:15.990917Z",657,{"id":1163,"title":1164,"summary":1165,"tags":1166,"translations":1171,"news_slug":1176,"published_at":1177,"created_at":1178,"image_url":27,"view_count":1179},"3583b78f-361b-4456-a4c1-962ff4275a8c","LobsterAI 上线多模态大模型矩阵：网易有道开源产品首次打通图片与视频生成","6月1日，网易有道宣布其开源产品 LobsterAI（龙虾）正式上线图片生成与视频生成能力，一次性接入 Seedream、Seedance、HappyHorse、MiniMax-Hailuo 等多个业界主流模型。这意味着网易有道成为国内首个将图片生成与视频生成能力整合进统一开源产品的大厂级玩家。LobsterAI 脱胎于网易有道的 AI 视觉技术积累，最初定位为图像生成方向的轻量开源工具。此次版本迭代后，产品架构扩展为「图片+视频双引擎」模式，支持多模型热切换——用户无需重新配置环境，即可在 Seedream 的高写实风格、Seedance 的艺术渲染能力、MiniMax-Hailuo 的视频动态效果之间自由切换，从而覆盖海报设计、短视频创作、多模态内容创作等不同场景。值得关注的是，LobsterAI 选择了一条「模型聚合」而非「自研一条龙」的道路。不同于 Midjourney 或 Sora 的封闭生态，LobsterAI 通过标准化的 API 接口层，将多个第三方模型打包为统一的调用入口。这本质上是一种模型路由（Model Routing）策略：让用户按需选型，而非被单一模型的能力边界所限制。从技术角度看，多模型聚合对底层的调度稳定性提出了更高要求。不同模型的输出格式、token 预算、延迟特性各有差异，能否在生产环境中保持一致的用户体验，是 LobsterAI 接下来需要证明的事。对于国内开源社区而言，LobsterAI 的路径提供了一种新思路：未必非要从零训练一个 SOTA 模型，整合与调优同样可以构建竞争壁垒。在开源模型生态日趋繁荣的当下，「连接者」的角色或许和「创造者」一样重要。",[1167,1168,1169,1170],{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},{"id":148,"name":149,"slug":149,"description":27,"color":27},[1172],{"id":1173,"lang":39,"title":1174,"summary":1175,"content":27},"9e88e39d-67e9-4252-87fd-e12c196fe941","LobsterAI launches an open multimodal model matrix","NetEase Youdao released LobsterAI on June 1, a multimodal large-model matrix that is the company's first open-source product to unify image and video generation. The matrix includes text-to-image, image-to-image, text-to-video, and image-to-video models, with a focus on the education and content creation use cases.","netease-lobsterai-image-video-model-routing","2026-06-01T05:00:00Z","2026-06-01T13:07:22.146578Z",389,{"id":1181,"title":1182,"summary":1183,"tags":1184,"translations":1189,"news_slug":1194,"published_at":1195,"created_at":1196,"image_url":27,"view_count":1197},"2bd17cfe-c17d-43d4-82ad-3688fa69fd3a","NVIDIA联手Black Forest Labs：FP8量化让FLUX.2进入RTX消费级显卡时代","Black Forest Labs 于 2026 年 5 月初发布了 FLUX.2 图像生成模型系列，拥有 320 亿参数，可在 ComfyUI 直接运行。然而，90GB VRAM 的需求让消费级 GPU 完全无法承载。NVIDIA 迅速介入，联合 Black Forest Labs 对 FLUX.2 进行 FP8 量化优化，成功将 VRAM 需求降低 40%，同时保持图像质量基本不变。配合 ComfyUI 的权重卸载（weight streaming）功能，RTX 消费级显卡现在也能运行这一旗舰模型，性能提升约 40%。\n\n这一合作背后有几个值得关注的技术信号。首先，FP8 量化正在成为大模型落地的标准路径——不是等下游厂商自己优化，而是上游芯片商主动介入，确保自家硬件不因内存墙被淘汰。其次，ComfyUI 作为开源社区枢纽，在模型—硬件的适配中扮演了关键角色，权重流式加载让显存和内存协同工作，绕过了单卡物理限制。对追求高分辨率生成的开发者而言，FLUX.2 + RTX 4080 以上显卡 + ComfyUI 的组合已进入实用阶段，不再只是实验室演示。",[1185,1186,1187,1188],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":29,"name":30,"slug":30,"description":27,"color":27},{"id":827,"name":828,"slug":828,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1190],{"id":1191,"lang":39,"title":1192,"summary":1193,"content":27},"799bbf3c-661c-43aa-9bbb-5989f734ec38","NVIDIA and BFL bring FP8-quantized FLUX.2 to consumer RTX","Black Forest Labs released the FLUX.2 image generation model family in early May 2026, with 32 billion parameters, runnable directly in ComfyUI. However, the 90GB VRAM requirement made it completely out of reach for consumer-grade GPUs. NVIDIA quickly stepped in, partnering with Black Forest Labs on FP8 quantization optimization for FLUX.2, successfully reducing VRAM requirements by 40% while keeping image quality essentially unchanged. Combined with ComfyUI's weight streaming feature, RTX consumer-grade GPUs can now also run this flagship model, with about 40% performance improvement.\n\nSeveral technical signals worth attention lie behind this partnership. First, FP8 quantization is becoming the standard path for large-model deployment — instead of waiting for downstream vendors to optimize themselves, upstream chip vendors actively step in to ensure their hardware isn't eliminated by the memory wall. Second, ComfyUI, as an open-source community hub, plays a key role in model-hardware adaptation, with weight streaming letting VRAM and RAM collaborate to bypass single-card physical limits. For developers pursuing high-resolution generation, the combination of FLUX.2 + RTX 4080 or above + ComfyUI has entered the practical stage, no longer just a lab demo.","nvidia-bfl-flux-2-fp8-rtx-40pct","2026-05-13T22:00:00Z","2026-05-13T22:07:29.247803Z",575,{"id":1199,"title":1200,"summary":1201,"tags":1202,"translations":1207,"news_slug":1212,"published_at":1213,"created_at":1214,"image_url":27,"view_count":1215},"1e0c52a9-c802-438d-9f21-4ea7ddc3d96a","Midjourney V8.1 发布：HD模式提速3倍，图像生成进入「高效量产」时代","4 月 30 日，Midjourney 正式发布 V8.1 版本，这是 V8 系列自 3 月上线以来的第三次迭代。相比 V8.0，V8.1 在渲染速度和成本控制上实现了质的飞跃：HD 模式提速 3 倍、成本降低 3 倍，并被设为默认模式；标准分辨率模式也提升了 50% 速度、降低 25% 成本，V8.1 标准模式 Full Quality 的出图速度已与 V7 的 Draft Mode 持平。\n\n**速度提升背后的技术逻辑**\n\nV8 系列基于全新 GPU 原生架构构建，V8.1 进一步针对渲染管线做了深度优化。HD 模式（输出 2K 分辨率）此前因计算量大、耗时长、成本高，让不少用户望而却步。此番将 HD 设为默认并大幅提速降价，意味着 Midjourney 正在将高质量图像生成推向「可大规模使用」的临界点。\n\n**社区功能回归与新工具**\n\nV8.1 还带回了此前在 V7 中大受好评的功能：Image Prompts（图像提示）和 Image Weights（图像权重）重新支持，用户可以将参考图融入提示词中精确控制生成结果。新增的 Prompt Shortener 则会在提示词超长时自动压缩优化，缓解了 V8 超长提示词容易失效的问题。Describe 功能也升级为更详细、更匹配 V8 风格的描述输出。\n\n**竞争格局观察**\n\n图像生成模型战场正在分化：Midjourney 专注速度和易用性，FLUX.2 走开源协作路线，GPT Image 2 依托语言模型优势做图文融合。V8.1 的提速降本策略，更像是在价格敏感场景中抢占身位。值得注意的是，V8 系列目前仍在 alpha 测试阶段，正式版的功能和定价尚未落定。\n\nMidjourney V8.1 的发布再次证明：图像生成模型的核心竞争力已从「能否生成」，转向「生成多快、成本多低」。2026 年的图像生成战场，速度和成本正在成为新的决胜手。",[1203,1204,1205,1206],{"id":483,"name":484,"slug":484,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1208],{"id":1209,"lang":39,"title":1210,"summary":1211,"content":27},"7eddfee8-8490-4af5-828a-d43fbf183634","Midjourney V8.1: HD mode three times faster","On April 30, Midjourney officially released version V8.1, the third iteration of the V8 series since its March launch. Compared to V8.0, V8.1 achieved a qualitative leap in rendering speed and cost control: HD mode is 3× faster with 3× cost reduction, and is now the default; standard-resolution mode also gets 50% speedup and 25% cost reduction, with V8.1 Standard Mode Full Quality's generation speed now matching V7's Draft Mode.\n\n**The technical logic behind the speed boost**\n\nThe V8 series is built on a brand-new GPU-native architecture, and V8.1 further optimizes the rendering pipeline in depth. HD mode (output 2K resolution) was previously so compute-intensive, time-consuming, and costly that many users were deterred. Setting HD as default and dramatically improving speed and pricing means Midjourney is pushing high-quality image generation to a \"mass-usable\" tipping point.\n\n**Community feature returns and new tools**\n\nV8.1 also brings back features that were well-received in V7: Image Prompts and Image Weights return, letting users blend reference images into prompts to precisely control generation results. The new Prompt Shortener automatically compresses and optimizes overly long prompts, addressing the V8 issue where overlong prompts easily fail. The Describe feature has also been upgraded to more detailed, V8-style output descriptions.\n\n**Competitive landscape observations**\n\nThe image-generation model battlefield is differentiating: Midjourney focuses on speed and ease-of-use, FLUX.2 takes the open-source collaboration route, GPT Image 2 leverages language-model advantages for image-text fusion. V8.1's speed-up and cost-down strategy looks more like grabbing position in price-sensitive scenarios. Notably, the V8 series is still in alpha testing, with the official version's features and pricing not yet finalized.\n\nMidjourney V8.1's release again proves: the core competitiveness of image generation models has shifted from \"can it generate\" to \"how fast and how cheap.\" In 2026's image generation battlefield, speed and cost are becoming the new deciding factors.","midjourney-v8-1-hd-3x-default","2026-05-12T05:15:00Z","2026-05-12T13:14:32.656911Z",476,{"id":1217,"title":1218,"summary":1219,"tags":1220,"translations":1225,"news_slug":1230,"published_at":1231,"created_at":1232,"image_url":27,"view_count":181},"fec5fd73-0995-4484-b7f2-0ec466dab080","Black Forest Labs 发布 FLUX.2：图像生成进入「分工协作」新时代","Black Forest Labs 正式发布 FLUX.2 图像生成模型家族。与以往「一个模型打天下」的思路不同，FLUX.2 带来了 Pro、Flex、Dev、Klein 四个变体，分别面向质量优先、商用可控、开源本地化、轻量实时四种不同场景需求。\n\n**技术层面**，FLUX.2 采用了与 FLUX.1 不同的架构策略，不再是单一的大模型，而是一组功能分化但共享底层的模型家族。这种「共享底层 + 任务专用头」的设计，类似 NLP 领域 LoRA 与 MoE 的思路组合——在保持核心能力的同时，降低了推理成本并提升了场景适配度。官方数据显示 FLUX.2 Pro 相比前代在图像质量评分上提升了 23%，而 Klein 变体则实现了单张图像生成在 1 秒以内完成。\n\n**实际影响**在于，FLUX.2 实际上在为「谁来用、怎么用」做了分工。专业设计师可以选 Pro 做品牌视觉；API 开发者可以选 Flex 做商业产品嵌入；开源社区则可以用 Dev\u002FKlein 在本地机器上跑。这种分层设计意味着，图像生成正式从「技术展示」走向「生产分工」。而生产分工往往是技术走向成熟的标志。\n\n**对行业的启示**：2024 到 2025 年间，图像生成模型的主旋律是「更逼真」；2026 年开始的主旋律正在变成「更实用」。FLUX.2 不是追求在所有指标上击败 FLUX.1，而是在说——同一个底座，可以长出不同用途的枝干。这比单纯堆参数更有工程意义。",[1221,1222,1223,1224],{"id":483,"name":484,"slug":484,"description":27,"color":27},{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1226],{"id":1227,"lang":39,"title":1228,"summary":1229,"content":27},"24a2cf44-aa2a-4d4d-affe-f02cd7a2f069","BFL ships FLUX.2: image generation enters divided labor","Black Forest Labs has officially released the FLUX.2 image generation model family. Unlike the previous \"one model to rule them all\" thinking, FLUX.2 brings four variants — Pro, Flex, Dev, Klein — targeting four different scenario needs: quality-first, commercial-controllable, open-source local, and lightweight real-time.\n\n**On the technical side**, FLUX.2 adopts a different architectural strategy from FLUX.1 — no longer a single large model, but a family of functionally differentiated models sharing a common foundation. This \"shared foundation + task-specific heads\" design is similar to the NLP-domain combination of LoRA and MoE thinking — while preserving core capability, it reduces inference cost and improves scenario fit. Official data shows FLUX.2 Pro improves image quality scores by 23% over the previous generation, while the Klein variant completes single-image generation in under 1 second.\n\n**The real impact** is that FLUX.2 actually divides \"who uses it, how they use it.\" Professional designers can choose Pro for brand visuals; API developers can choose Flex for commercial product embedding; the open-source community can use Dev\u002FKlein to run on local machines. This tiered design means image generation is officially moving from \"tech showcase\" to \"production division of labor.\" And production division of labor is often the sign of technology maturing.\n\n**Industry implications**: The main theme of image generation models from 2024 to 2025 was \"more realistic\"; the main theme from 2026 onward is becoming \"more practical.\" FLUX.2 isn't pursuing beating FLUX.1 on all metrics — it's saying that from the same foundation, different-purpose branches can grow. This is more engineering-meaningful than simply stacking parameters.","flux-2-pro-flex-dev-klein-4-variants","2026-05-08T04:00:00Z","2026-05-08T04:07:17.453970Z",{"id":1234,"title":1235,"summary":1236,"tags":1237,"translations":1241,"news_slug":1246,"published_at":1247,"created_at":1248,"image_url":27,"view_count":1249},"69f2bd5f-9e25-4a6f-b0da-b94bd40ceeb6","图像模型成AI应用增长新引擎：ChatGPT单月创收7000万美元","最新数据显示，图像生成模型已成为AI移动应用增长的核心驱动力。据Appfigures报告，图像模型更新带来的下载量增幅是传统聊天模型升级的6.5倍，这一趋势在头部应用中表现尤为突出。Gemini在去年8月推出图像模型后，28天内新增下载超2200万次，环比激增4倍以上。ChatGPT的GPT-4o图像功能更为惊人——单次发布便带来超1200万增量安装，约为此前GPT-4o、GPT-4.5及GPT-5系列发布总和的4.5倍。更有商业价值的是，下载转化付费的效率差异显著。ChatGPT凭借GPT-4o图像模型，28天内创造了约7000万美元的消费者支出，而Gemini的Nano Banana同期仅带来约18.1万美元。这背后的差异在于产品成熟度与用户付费心智的建立。本质上看，图像生成能力已成为AI产品的钩子——用它吸引用户首次尝鲜，但真正的商业闭环仍需依托对话智能与工作流整合。这给开发者的启示是：纯视觉噱头不够，必须将图像能力嵌入真实使用场景才能完成付费转化。",[1238,1239,1240],{"id":53,"name":54,"slug":54,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1242],{"id":1243,"lang":39,"title":1244,"summary":1245,"content":27},"bb8c3ffb-aeb5-4435-a387-88f6a6e4a0fb","Image models drive growth: ChatGPT earns $70M in a month","Latest data shows that image generation models have become a core driver of AI mobile app growth. According to Appfigures, the download increase from image-model updates is 6.5× that of traditional chatbot model upgrades, a trend especially pronounced in top apps. After Gemini launched its image model last August, it gained over 22 million new downloads in 28 days, a 4×+ month-over-month surge. ChatGPT's GPT-4o image feature is even more stunning — a single release brought over 12 million incremental installs, about 4.5× the combined total of previous GPT-4o, GPT-4.5, and GPT-5 series releases. More commercially valuable is the significant difference in download-to-pay conversion. ChatGPT, leveraging the GPT-4o image model, generated about $70 million in consumer spending in 28 days, while Gemini's Nano Banana only brought about $181,000 in the same period. The difference behind this lies in product maturity and the establishment of user pay-mindset. In essence, image generation capability has become the AI product's hook — using it to attract users' first taste, but the real business loop still relies on conversational intelligence and workflow integration. The lesson for developers: pure visual gimmick isn't enough, image capability must be embedded in real usage scenarios to complete pay conversion.","chatgpt-image-revenue-70m-2026","2026-05-05T04:00:00Z","2026-05-05T04:05:30.339956Z",219,{"id":1251,"title":1252,"summary":1253,"tags":1254,"translations":1259,"news_slug":1264,"published_at":1265,"created_at":1266,"image_url":27,"view_count":1267},"18df2ab9-affd-4ba8-adaa-03cc7e3a2317","GPT Image 2 发布：OpenAI 首次将推理能力注入图像生成","4月21日，OpenAI 正式发布 GPT Image 2（ChatGPT Images 2.0），这是 DALL-E 3 的继任者，也是业界首款真正将 O 系列推理能力融入图像生成的模型。与传统扩散模型不同，GPT Image 2 在渲染像素之前，会先研究提示中的实体关系、规划图像布局、推理细节约束，然后才输出——这意味着它真正先思考再作画。\n核心技术突破体现在三个方面。第一，Agentic 架构：模型不再遵循提示→直接渲染的直线路径，而是加入了研究、规划、验证四阶段流程，显著提升了杂志排版、多格漫画、复杂信息图等场景的一次成功率。第二，多语言文字渲染：支持拉丁、日韩 CJK、印地语、孟加拉语等，字符级准确率达 99%，解决了图像模型长期难以准确渲染文字的顽疾。第三，原生 2K（2048 像素）分辨率输出，满足商业印刷级别需求。\nbenchmark 方面，GPT Image 2 以 +242 分优势登顶 Image Arena 排行榜。背后由 GPT-5.4 主干网络驱动。API 定价：图像输入 tokens 8美元\u002F百万（缓存命中仅2美元），输出 tokens 30美元\u002F百万。ChatGPT 和 Codex 用户4月22日起全面开放，开发者 API 预计5月上线。\nGPT Image 2 的意义不仅是又一款高质量图像模型，更重要的是它重新定义了多模态生成的竞争逻辑。当业界还在优化扩散模型的步数与调度时，OpenAI 已将竞争焦点从生成质量转向生成可靠性——让复杂 prompt 不再翻车，才是下一代创作工具的核心壁垒。这一路线一旦被验证跟进，多模态模型的 Agentic 化将成为下一阶段的主战场。",[1255,1256,1257,1258],{"id":189,"name":190,"slug":190,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":101,"name":102,"slug":102,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1260],{"id":1261,"lang":39,"title":1262,"summary":1263,"content":27},"0e67c2ab-20df-4b78-8296-b8cbc851ea62","GPT Image 2: OpenAI brings reasoning into image generation","On April 21, OpenAI officially released GPT Image 2 (ChatGPT Images 2.0), the successor to DALL-E 3 and the industry's first model that truly integrates O-series reasoning capability into image generation. Unlike traditional diffusion models, GPT Image 2 first studies entity relationships in the prompt, plans image layout, reasons about detail constraints, before outputting — meaning it truly thinks first, then paints.\n\nCore technical breakthroughs manifest in three aspects. First, Agentic architecture: the model no longer follows a straight line from prompt to direct rendering, but introduces four stages of research, planning, and verification, significantly improving first-pass success rates for magazine layouts, multi-panel comics, and complex infographics. Second, multilingual text rendering: supports Latin, Japanese\u002FKorean CJK, Hindi, Bengali, etc., with character-level accuracy reaching 99%, solving the long-standing problem of image models being unable to render text accurately. Third, native 2K (2048 pixel) resolution output, meeting commercial print-level requirements.\n\nOn benchmarks, GPT Image 2 tops the Image Arena leaderboard with a +242 score advantage. Driven by the GPT-5.4 backbone network. API pricing: $8\u002Fmillion image input tokens ($2 with cache hit), $30\u002Fmillion output tokens. ChatGPT and Codex users gain full access from April 22, developer API expected in May.\n\nGPT Image 2's significance isn't just being another high-quality image model, but more importantly that it redefines multimodal generation's competitive logic. While the industry is still optimizing diffusion model steps and scheduling, OpenAI has shifted the competitive focus from generation quality to generation reliability — making complex prompts stop failing is the core moat of the next generation of creative tools. Once this path is validated and followed, multimodal models' Agentic-ization will become the next main battlefield.","gpt-image-2-openai-agentic-2k-reasoning","2026-04-28T10:00:00Z","2026-04-28T10:08:34.914772Z",532,{"id":1269,"title":1270,"summary":1271,"tags":1272,"translations":1278,"news_slug":1283,"published_at":1284,"created_at":1285,"image_url":27,"view_count":1286},"c4976de5-150e-4dc7-bf86-68d7d5f81be1","分文本条件增强：DiT-ST革新文本到图像扩散模型","近日，同济大学等机构的研究者在arXiv发表论文，提出了一种名为DiT-ST的创新框架，通过分文本条件（Split-Text Conditioning）显著提升了扩散Transformer在文本到图像生成中的表现。\n\n**技术突破点**：\n传统文本到图像扩散模型存在完整文本理解缺陷，由于文本长度限制、softmax竞争和位置偏差等问题，导致属性绑定错误、语义混乱等问题。DiT-ST通过三个关键创新解决了这一挑战：\n\n1. **文本解析**：利用大型语言模型（LLM）将复杂文本解析为语义基元（对象、关系、属性）并构建层次化图谱\n2. **层次化输入**：将完整文本转换为简化的分文本输入，降低语法复杂度\n3. **增量注入**：根据去噪阶段对不同语义类型的敏感性，按对象-关系-属性的优先级顺序在不同时间步注入语义信息\n\n**性能提升**：\n- 在GenEval基准测试上达到69%的整体准确率，接近SDv3.5 Large（71%）\n- COCO-5K数据集CLIPScore达34.09，超越SDv3.5 Large约4.1%\n- 对复杂长文本表现更加鲁棒，解决了现有模型对文本长度的敏感性\n\n这项研究不仅提升了生成质量，更重要的是揭示了扩散模型在不同去噪阶段对不同语义类型敏感性的内在规律，为理解扩散模型的语义建立过程提供了新视角。这种基于语义分解的方法有望推动多模态生成技术的进一步发展。",[1273,1274,1275,1276,1277],{"id":25,"name":26,"slug":26,"description":27,"color":27},{"id":167,"name":168,"slug":168,"description":27,"color":27},{"id":32,"name":33,"slug":33,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},{"id":58,"name":59,"slug":59,"description":27,"color":27},[1279],{"id":1280,"lang":39,"title":1281,"summary":1282,"content":27},"8e720024-4394-40f5-87c8-777e1b54cd09","DiT-ST: partitioned text conditioning reshapes T2I diffusion","Researchers from Tongji University and other institutions have published a paper on arXiv proposing DiT-ST, an innovative framework that uses Split-Text Conditioning to significantly improve diffusion Transformers' performance on text-to-image generation.\n\n**Technical Breakthroughs:**\nTraditional text-to-image diffusion models suffer from incomplete-text understanding flaws — text length limits, softmax competition, and positional bias lead to attribute-binding errors and semantic confusion. DiT-ST addresses these challenges through three key innovations:\n\n1. **Text Parsing:** Uses a large language model (LLM) to parse complex text into semantic primitives (objects, relations, attributes) and build a hierarchical graph.\n2. **Hierarchical Input:** Converts complete text into simplified split-text inputs, reducing grammatical complexity.\n3. **Incremental Injection:** Based on the denoising stage's sensitivity to different semantic types, injects semantic information at different timesteps in an object-relation-attribute priority order.\n\n**Performance Gains:**\n- 69% overall accuracy on the GenEval benchmark, close to SDv3.5 Large (71%).\n- CLIPScore of 34.09 on the COCO-5K dataset, about 4.1% above SDv3.5 Large.\n- More robust on complex long text, solving existing models' text-length sensitivity.\n\nThis work not only improves generation quality — more importantly, it reveals diffusion models' intrinsic rules of sensitivity to different semantic types at different denoising stages, offering a new perspective on understanding how diffusion models semantically build up. This semantic-decomposition-based approach is expected to push multimodal generation further.","dit-st-split-text-conditioning-tongji-69pct-geneval","2026-04-21T07:00:00Z","2026-04-21T07:05:56.050728Z",415,{"id":1288,"title":1289,"summary":1290,"tags":1291,"translations":1295,"news_slug":1300,"published_at":1301,"created_at":1302,"image_url":27,"view_count":1303},"3644ee0e-e0fd-491b-afef-5f269c91f0f2","阿里发布 Wan2.7-Image：统一图像生成与编辑模型，文字渲染和多图控制是杀手锏","阿里通义实验室发布 Wan2.7-Image，一个统一了图像生成与编辑的全链路模型。与以往文生图和图像编辑分属不同模型的做法不同，Wan2.7-Image 将文生图、图生图、指令编辑、交互式编辑整合在同一个架构中，用户无需在多个工具间切换。\n\n在人类偏好盲测中，Wan2.7-Image 的文生图能力超过了 GPT-Image 1.5 和国内主流模型，文字渲染、写实成像和世界知识等指标接近 Google Nano Banana Pro。\n\n六项核心能力中，有两项尤其值得关注。第一是**精细人物控制**：支持脸型、眼型、骨相等面部特征的自定义，试图解决 AI 人像长期存在的\"千人一面\"问题。第二是**文字渲染**：支持 4000+ 字符生成，覆盖中文简繁体、英日韩德法等多语言，大幅降低了 AI 海报、电商素材制作中的文字乱码问题。\n\n此外，Wan2.7-Image 的\"调色盘\"功能支持通过 Hex 色值精确控制画面色彩，对品牌设计和产品图制作非常实用。交互式编辑模式允许用户通过自然语言指令对生成图像进行局部修改，降低了非专业用户的使用门槛。\n\n值得注意的是，Wan2.7 不仅有图像模型，视频模型升级也在推进中。阿里正在构建的不只是单一生成工具，而是一个从图像到视频的多模态生成平台。在图像生成领域，从 Midjourney 到 FLUX 再到 Wan2.7-Image，竞争焦点正在从\"画得好看\"转向\"画得可控\"——这是一个更难也更有价值的方向。",[1292,1293,1294],{"id":167,"name":168,"slug":168,"description":27,"color":27},{"id":77,"name":78,"slug":78,"description":27,"color":27},{"id":35,"name":5,"slug":5,"description":27,"color":27},[1296],{"id":1297,"lang":39,"title":1298,"summary":1299,"content":27},"2f911aa9-f815-4895-a95a-e46cc1765a2c","Wan2.7-Image: unified editing with killer text rendering","Alibaba's Tongyi Lab has released Wan2.7-Image, a full-pipeline model that unifies image generation and editing. Unlike the traditional split between text-to-image and image-editing models, Wan2.7-Image integrates text-to-image, image-to-image, instruction editing, and interactive editing into a single architecture — users no longer need to switch between multiple tools.\n\nIn human-preference blind tests, Wan2.7-Image's text-to-image capability surpassed GPT-Image 1.5 and major domestic models, with text rendering, realistic imaging, and world-knowledge metrics approaching Google Nano Banana Pro.\n\nOf the six core capabilities, two are especially worth highlighting. First is **fine-grained character control**: support for customizing facial features (face shape, eye shape, bone structure) — an attempt to solve the long-standing \"thousand faces look the same\" problem in AI portraits. Second is **text rendering**: supports 4000+ character generation, covering simplified and traditional Chinese, English, Japanese, Korean, German, French and more, dramatically reducing text-garbling issues in AI posters and e-commerce material production.\n\nWan2.7-Image's \"palette\" feature also supports precise color control via Hex values, which is very practical for brand design and product imagery. The interactive editing mode allows users to make local modifications to generated images via natural-language instructions, lowering the bar for non-professional users.\n\nNotably, Wan2.7 is not just an image model — video model upgrades are also in progress. What Alibaba is building is not a single generation tool, but a multimodal generation platform from image to video. In image generation, from Midjourney to FLUX to Wan2.7-Image, the competitive focus is shifting from \"looking good\" to \"controllable\" — a harder and more valuable direction.","alibaba-wan-2-7-image-unified-text-render-4000-chars","2026-04-17T12:42:00Z","2026-04-17T12:43:20.969226Z",289]