[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-asuka-fm-single-rtx-4090-t2i-pretraining":3,"topics-all":38,"news-related-5a1cb663-bfe6-4c9a-bc4e-e1c8e08f942f":66},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"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.","55 美元,在今天能买到什么?几天的前沿文生图 API 额度,或者一顿不错的晚餐。西班牙 Universidad Carlos III de Madrid 的四人学术团队给出了第三个答案: 按论文自报的口径,用这笔钱租一张 RTX 4090 跑约 160 GPU 小时,从零预训练出一个 206M 参数的文生图 DiT 基座 —— 是从零,不是微调谁家的权重(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26437)。\n\n## 从零训一个文生图模型,难在哪\n\n论文在引言里把对手摆得很清楚: FLUX.1 12B, SD3 8B, Qwen-Image 20B,前沿文生图全部在工业集群里滚参数。学术组想自己从零训一个 DiT,会同时撞上三堵墙: 注意力对序列长度是二次方成本;真实图片宽高比五花八门,分桶喂进训练会把 torch.compile 的 JIT 编译图反复打断,重编译吃掉大量步时;更隐蔽的是冻结 VAE 这堵墙 —— 论文算过一笔账,预先把几十万张图的 32 通道 float16 latents 全部落盘需要数 TB 高速存储,消费级 NVMe 装不下,只能边训边编码。论文实测在 4090 上,冻结 VAE 编码每步吃掉 164.92 ms,占 62.0% 的步时延,而 DiT 前向加反向合计不到 70 ms。到了 1024² 分辨率,on-the-fly VAE 编码更是占到 72.6% 的墙钟时间。\n\n## 四件套怎么拆墙\n\n第一件,把 AdaLN 调制从每个 block 里拆掉,时间条件改成前置的连续正弦时间 token,backbone 参数省下 28%,论文称生成质量不受影响。第二件,序列稀疏化 SPRINT —— 这是引用 Park 等人的先前工作(https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.21986),本文把它扩展到开放域多模态条件: 75% 稀疏的 token 序列穿过 9 个重的中间块,注意力 FLOPs 砍掉一半以上,512² 分辨率实测吞吐 108.5 img\u002Fs 对 64.8 img\u002Fs。第三件,连续宽高归一 RoPE 配平移方形裁剪,张量形状彻底静态化,重编译消失,还白送了零样本虚拟相机控制。第四件,一个 10M 参数的 Patch Detailer Head 用 v-prediction 把潜空间权重迁到像素空间 —— 推理端 VAE 解码在少步数场景下占 70% 以上时延,像素直出把这堵墙整个绕开。\n\n训练端的 VAE 墙另有解法: 256² 下全部 62.5 万张图的 latents 约 52GB,直接塞进 128GB DDR5 内存,磁盘 IO 和自编码器时延一起消失,吞吐到 282.6 img\u002Fs,实测 MFU 86.8%;到 512²,latents 膨胀到约 203GB,内存装不下,改走网络流式读取。整套系统跑在 Ryzen 9 7950X + 128GB 内存 + 单张 4090 上,平均物理 MFU 75.6%。\n\n## 成绩单,以及该泼的冷水\n\n论文实测: 256px 阶段(68 小时)GenEval 0.588 \u002F DPG 78.96;512px 阶段(81 小时)GenEval 0.609 \u002F DPG 78.51,全程无基准向微调,256px 的 DPG 反而比 512px 高出一截。对照表里,SD3 Medium(2B,集群)是 0.620 \u002F 84.08;学术基线 MiniT2I-L\u002F16(0.91B,8×H100 连跑 27 天)是 0.883 \u002F 84.80 —— 但论文提醒,后者在 BLIP3o-60K 上微调过,而该数据集与 GenEval 句式直接对齐,分数天然占便宜。论文自称在效率-质量的帕累托前沿上立了新标准,这句是自称。\n\n冷水也得泼足。绝对分数仍差一截,这是 62.5 万张训练图对上亿张图的差距,论文自己承认复杂人群场景和长尾概念仍是更大的模型更强;多词文字渲染不稳定;像素空间下采样步数一多(N≥13)延迟反而高于潜空间。还有一笔对不上的账: 摘要写 160 GPU 小时(按论文表注 0.33 美元\u002F卡时折算即 55 美元以内),但训练路线图三个阶段分别是 68h, 81h, 57.7h,加总约 207 小时 —— 「55 美元」是论文摘要的自报口径,引用时请带着这个心眼。\n\n## 不是孤例,是一条路线\n\n「小算力从零训扩散」这条路线已有独立佐证: 单人作者 Yeh 的 HDM 靠一个 U 形 transformer(XUT),用四张 RTX 5090 从零训出 1024² 文生图模型,自报成本 535-620 美元(https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.06068);Asuka-FM 的参考文献里还引了 Sehwag 等人 CVPR 2025 的 micro-budget 扩散训练。区别在完成度: Asuka-FM 把数据(62.5 万张开放多源 + 33.7 万张精选 Danbooru),代码(GitHub: deeplearningcafe\u002Fasuka-fm)和权重(HuggingFace: aipracticecafe\u002Fasuka-fm)整套开放,并把这条路线推到了开放域。\n\n所以呢。对个人和小实验室,「训一个自己的文生图基座」从口号变成了 55 美元级的周末工程(论文口径);论文展望里格外值钱的一句是: RL 后训练时每个 prompt 的 12 条 rollout 意味着 12 次 VAE 解码,占掉 60% 以上显存,像素直出让消费级工作站上的在线 RL 变得可行。基座不再必然属于数据中心 —— 至少在 200M 这个量级,这句话现在有账可算了。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26437","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c883fd20-1d66-4fb7-9fc7-320fa7f87023","text-to-image",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"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","2026-11-21T13:28:16.722831Z",true,"agent",35,[39,48,57],{"slug":40,"tag_slug":23,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"image-generation","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","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":58,"tag_slug":58,"title_zh":59,"title_en":60,"intro_zh":61,"intro_en":62,"id":63,"is_active":35,"created_at":64,"modified_at":65},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":67},[68,73,78,83,88,93],{"id":69,"title":70,"news_slug":71,"published_at":72},"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","moebius-0-2b-tiny-specialist-inpainting","2026-06-17T15:35:38+00:00",{"id":74,"title":75,"news_slug":76,"published_at":77},"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","linum-pjit-pixel-space-text-to-image-21k","2026-11-07T03:08:12+00:00",{"id":79,"title":80,"news_slug":81,"published_at":82},"20f285d4-4a2f-4722-973a-465e3da4c5e1","蚂蚁IDM蒸馏:2步出图,长文本渲染94.25%","idm-intrinsic-distribution-matching-longtext","2026-10-30T11:01:03+00:00",{"id":84,"title":85,"news_slug":86,"published_at":87},"3bf416bb-a989-4b02-8341-dbafc7c96c48","MOSS-Image 开源:5B 双语生图,文字排版可编程","moss-image-bilingual-text-rendering-base","2026-10-18T19:08:44+00:00",{"id":89,"title":90,"news_slug":91,"published_at":92},"b88d9e4c-1450-4349-a3cd-c2f8c0ce50eb","FLUX 3 Klein 开源:5.6B,8GB 显存能跑","flux-3-image-klein-open-weights","2026-10-17T13:20:00+00:00",{"id":94,"title":95,"news_slug":96,"published_at":97},"0b0af156-7b5c-47d8-bfc2-de8cc0b22ce2","混元开源Prism:稀疏注意力让2K音视频训练提速2.5倍","prism-2k-video-audio-sparse-attention","2026-10-13T23:13:00+00:00"]