[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-linum-pjit-pixel-space-text-to-image-21k":3,"topics-all":38,"news-related-3f4ca608-fb14-45b2-aeed-122264f141e7":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},"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。","图像生成这一轮，大家习惯谈参数量和跑分。11 月 4 日，一个只有两个人的小实验室 Linum 公开了另一笔账：预训练加后训练烧掉 8,415 个 H100 小时，官方折算总成本 21,038 美元。权重和代码按 Apache 2.0 开源，连完整训练配方（样本数、批量、优化器分工、噪声调度、学习率曲线）一起贴在博客上（原始博客：https:\u002F\u002Fwww.linum.ai\u002Ffield-notes\u002Fpjit-21k）。\n\n比数字更值得看的是它背后的架构选择。\n\n## 拆掉 VAE，直接在像素上做扩散\n\n主流文生图几乎都在潜空间跑：先用 VAE 把图像压成低维潜变量，再在潜空间做扩散。Linum 判断这条路有经验上限——常见的 16×16 令牌压缩基本到顶（他们拿 FLUX-2 举例说明主流做法都是 16×16），而把 VAE 的重建质量做上去，在下游生成里反而可能更差。\n\n他们于是把 VAE 整个拆掉，让扩散模型直接在像素上跑，自己学一套专为生成定制的潜空间。代价是单令牌携带的信息更多，patch 尺寸就从 16×16 提到 32×32：单块 32×32×3 = 3,072 维 RGB 压进 1,024 维，压缩率 3:1，注意力上下文长度缩到 16×16 方案的四分之一。\n\n模型叫 Pyramid-JiT（P-JiT）：主干 22 个 DiT 块，前面各有 2 个图像 refiner 和 2 个文本 refiner 块，主干隐藏维度 2,944、23 个注意力头。文本侧用冻结的 Qwen3.5-4B，在第 7、15、27 层取出表示拼接。采样时 2.18B 参数，加上训练期专用的读出头约 0.24B，训练时约 2.42B。采样默认 50 步 Euler、自适应投影引导强度 15，权重 fp32 8.7 GB，文本编码器 bf16 8 GB。\n\n## 多分辨率读出：让主干自己打草稿\n\n另一处设计是「读出」（readout）。因为模型用 x-prediction 在像素空间训练，主干里任何一块的输出其实都是最终图像的粗糙版本。Linum 在第 10、16 块各挂一个线性头，分别预测四分之一和一半分辨率的干净图像，给早期层补梯度信号；第 5 块还有 PixelREPA 读出，把隐藏状态对齐到 DINOv3-L 特征。\n\n10 月 6 日那篇技术博客给出的对照（同一套 22 块主干、同一配方跑 75M 样本）：单读出基线的 512×512 验证损失 0.1037，换成 128²\u002F256²\u002F512² 三级读出降到 0.0934，降幅 10.0%；全分辨率三级读出 0.0948（降幅 8.5%），给低分辨率读出降权则 0.0983（降幅 5.1%）。同一篇文章称 P-JiT 追平上一代 Linum v2 的 FD-DINOv2 分数，用的训练样本少了 11.3 倍（https:\u002F\u002Fwww.linum.ai\u002Ffield-notes\u002Fpyramid-jit）。\n\n## 这份账单长什么样\n\n预训练 20,348 美元 = 8,139 H100 小时 × 每小时 2.50 美元；后训练 690 美元 = 276 H100 小时；合计 21,038 美元、8,415 H100 小时。硬件是 64 张 H100（8 个节点），bf16 加 FSDP2，每卡 32 张图。样本分两段：512² 阶段 610.27M（297,985 步），切到 1024² 后 49.97M（24,400 步），全局批量都是 2,048。\n\n优化器混着用：Muon（动量 0.95、带 Nesterov）作用在 2D block 矩阵上，其余参数用 AdamW，Muon 的更新按 RMS 对齐到 AdamW。学习率峰值 2e-4，线性 warmup 走 1M 样本，平到 550M，最后 60M 用 1−√ 衰减到 2.5e-5。噪声调度走 curriculum：前 80M 样本用 LogitNormal(0.8, 0.8) 把分布压在高噪区先学结构，之后换成 LogitNormal(−0.2, 1.0) 拉开分布学细节；切到 1024² 时噪声尺度从 2 翻到 4，用 RoPE 插值对齐坐标轴。后训练用 FDFO（Finite Difference Flow Optimization）的一个变体，配 JEV-Image 和人机协同打分。\n\n## 短板，官方自己写清楚了\n\n模型完全不会在图里写字，也做不了海报——这类文本密集数据在预处理时被整块过滤，官方说理由是他们的目标是动画工具，不需要这类数据。此外全身照、多实体（六辆车、成群的人）、多人交互、世界知识（吉他六根弦、骰子点数）都是明确写下的边界。仓库里贴的失败样例里，腿会被拉长、手会多出一只、骰子每个面都是六点。\n\n跑分上它也不是最强的。同一批模型的 GenEval（原始 prompt）：Linum v2 0.561、JiT-DDT 0.641、P-JiT 0.612；换成官方自己写的稠密 caption 后是 0.855 \u002F 0.808 \u002F 0.817，自家上一代反超。DPG-Bench 上 P-JiT 是 84.05 和 84.31，Linum v2 是 84.79 和 84.54。他们的解释是：像素空间模型出图更贴近底层图像分布，但在计数这类 prompt following 上像是训练不足——放大 prompt 后，计数任务上 Linum v2 是 0.80，P-JiT 只有 0.56。\n\n## 为什么这份账单值得单独看\n\n一是配方可抄。不止权重和推理代码，噪声调度怎么排、Muon 与 AdamW 怎么分工、梯度范数超滚动中位数 10 倍时暂停 EMA 250 步（所谓 spike-aware EMA），这类细节通常都被藏起来。\n\n二是它不是终点。官方写得很直白：这是通往下一款视频模型的垫脚石，因为视频模型几乎都从图像模型长出来，得先学名词再学动词。Linum v2 生成 5 秒 720p 视频要在 110K 个令牌上算注意力，而 LLM 有 97% 的训练阶段上下文不超过 8K——这就是他们要去掉的成本。社区也按研究件看它：HN 上 95 分 9 条评论，有人问展示样本挑了多少，有人指出它仍有文生图常见毛病——过度平滑、塑料感、肢体扭曲。模型卡和 GitHub README 由 Claude Opus 5.5 写成，模型、训练和代码审阅归 Linum 自己。\n\n所以呢？这 21,038 美元不是重点，重点是它把「训一个能用的文生图模型」的账单和配方同时摆上台面，让「自己训一个」从玄学变成能照抄的工程问题。下一个问题更有意思：当这套像素空间配方真的搬到视频上，那笔省下来的注意力开销，够不够把动画工具的价格打下来一个数量级。","https:\u002F\u002Fwww.linum.ai\u002Ffield-notes\u002Fpjit-21k","3c1def59-8c1e-46ec-9501-1c9801a6a648",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"035f9f3c-7983-44b1-aae0-3af221879109","en","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","2026-11-07T03:08:18.790906Z",true,"agent",255,[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},"5a1cb663-bfe6-4c9a-bc4e-e1c8e08f942f","单张4090从零训文生图:论文自报160卡时55美元","asuka-fm-single-rtx-4090-t2i-pretraining","2026-11-21T13:45:00+00:00",{"id":74,"title":75,"news_slug":76,"published_at":77},"bbbda7ff-b1c7-4142-b37f-2b3fd0547e02","Fibo 2.0 开源:生成编辑合一,登顶开源编辑榜","bria-fibo-2-0-unified-editing-generation","2026-10-21T21:08:11+00:00",{"id":79,"title":80,"news_slug":81,"published_at":82},"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":84,"title":85,"news_slug":86,"published_at":87},"f8725ca7-b0b9-430f-9c63-dabe9a2c1d57","Beam 501B官宣:23B激活参数打效率牌","reflection-beam-501b-open-weights","2026-10-06T13:13:51+00:00",{"id":89,"title":90,"news_slug":91,"published_at":92},"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","qwen-image-2-1-open-source-7b","2026-09-20T16:00:00+00:00",{"id":94,"title":95,"news_slug":96,"published_at":97},"30fca629-bace-4832-9789-b44aa8c8989d","学生团队从零训出开源 7B 模型 ZGCM-1:数学推理硬刚 235B 前沿","zgcm-1-open-7b-foundation-model","2026-09-15T19:10:00+00:00"]