[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-idm-intrinsic-distribution-matching-longtext":3,"topics-all":38,"news-related-20f285d4-4a2f-4722-973a-465e3da4c5e1":57},{"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},"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。","文生图模型的推理账单,几乎全部来自一件事:网络评估次数。扩散模型和流匹配模型靠反复去噪把噪声变成图像,每多一次前向就多一份延迟。一步生成和几步生成把这条曲线压平,代价是采样过程中几乎没有纠错机会——提示词里要求渲染的每一个字符,都得在仅剩的一两次前向里一次写对。\n\n长文本渲染把这道题摆到了台面上。海报、菜单、书封上的成串中英文,既得排版协调,又得字字正确。这正是蚂蚁(inclusionAI)这篇新论文挑的战场,论文原文见 https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39633 。\n\n## 核心思路:让生成器自己给自己当裁判\n\n论文的做法叫 Intrinsic Distribution Matching(IDM),落点在分布匹配蒸馏(DMD)这条线上。DMD 用分数差把生成分布往参考分布上拉,分数差来自一个预训练教师给出的参考分数,减去一个辅助假分数网络估计的当前生成分数。IDM 沿用前人做法,用内部的 velocity 分支来估计这个差,省掉了单独跑一个假分数网络的成本。\n\n新增的部分是显式的矩约束。论文把 reverse-KL 场拆成高斯矩分量和非高斯残差,由此指出有一批统计量是可以被直接约束的。IDM 在保留完整 DMD 修正的同时,去匹配特征均值和逐坐标标准差——而这些特征取自生成器自己的指数滑动平均(EMA)副本,不需要任何外部特征编码器。\n\n这一点是整篇论文最值得注意的地方。过去的做法是外挂一个 DINOv2 之类的特征编码器来提供\"锚点\",多一个模型就多一份显存、多一段工程复杂度。IDM 的答案是:生成器自己的 EMA 副本已经带了足够的结构信息,拿它当锚点就行。\n\n## 数字:长文本分数确实拉开了差距\n\n论文报告在 Z-Image-Turbo 6B 上,2 次网络评估(2 NFE)时,LongText-CN 达到 94.25%,LongText-EN 达到 90.20%,是论文对比的这一批方法里同预算下的最高分。对照组里,Z-Image-Turbo++ 是 91.62 和 89.88,APEX 是 90.38 和 88.76。换成百分点,IDM 比 Z-Image-Turbo++ 高出 2.63 和 0.32,比 APEX 高出 3.87 和 1.44。\n\n不过这张表有两个诚实的限定,论文自己写得很清楚。一是 8 NFE 的 Z-Image-Turbo 参考值在 LongText-EN 上仍然更高(91.74),也就是英文侧还没被完全追平。二是 APEX 的 DPG-Bench 是 88.64,高于 IDM 的 86.82。另外论文明确指出,这些公开参考模型的训练设置和数据各不相同,只有在同一套配置下的消融实验才真正可比。\n\n消融表是这篇论文最有说服力的部分。同样是 Z-Image 全参数、2 NFE:\n\n- 去掉 DMD 也不加特征矩损失:OneIG 52.25 \u002F GenEval 76.24 \u002F LongText-CN 92.70\n- 只保留内部 DMD、不加矩损失:52.87 \u002F 77.46 \u002F 93.87\n- 改用 DINOv2 特征做矩匹配:53.00 \u002F 77.75 \u002F 94.13\n- IDM,改用生成器 EMA 特征:53.06 \u002F 77.81 \u002F 94.25\n\n在内部 DMD 对照之上加 EMA 特征矩损失,五项指标分别带来 +0.19 OneIG、+0.35 GenEval、+0.28 DPG-Bench、+0.38 LongText-CN、+0.41 LongText-EN 的提升。有意思的是,生成器自己的 EMA 特征不但没有明显输给外接的 DINOv2,在多数项上还略高一点——这才是\"内生\"这个标题真正的证据。\n\n速度数据同样漂亮。全参数 Z-Image 在单张 A100、BF16 下,1 NFE 延迟 0.25 秒、2 NFE 延迟 0.30 秒;Qwen-Image 20B 分别是 0.35 秒和 0.44 秒。吞吐方面,Z-Image 2 NFE 能到 5.18 张\u002F秒,Qwen-Image 2 NFE 是 3.28 张\u002F秒。rank-64 的 LoRA 改动 0.4B 参数,基本保住了全参数模型的 GenEval 表现。训练侧的成本也交代了:Z-Image 训 15K 步用 32 张 H100 80GB,Qwen-Image 的 LoRA 训 10K 步用 8 张 A100 80GB,全局 batch size 64。\n\n## 泼点冷水:权重还没放出来\n\n这是必须点明的一点。GitHub 仓库(https:\u002F\u002Fgithub.com\u002FinclusionAI\u002FIDM)顶端的徽章写得很直接——\"Model weights: Coming soon\",Checkpoints 一节里 Z-Image-Turbo 6B 和 Qwen-Image 20B 的全参数与 r=32\u002Fr=64 LoRA 权重全部标着 \"Coming soon\"。仓库 README 里还有一句自我约束的说明:\"以下所有数值均为论文报告的结果,并非运行本仓库的实测结果。\"\n\n换句话说,目前跑不了、复现不了。表格里的分数是自报口径,第三方独立复现尚不存在。这不影响论文本身的方法学价值,但对想立刻上手的人是个明确的等待信号。README 也提醒,基座模型权重或 APEX 权重都不是 IDM checkpoint,拿它们复现不出上面的结果。\n\n训练数据规模也是一个变量。参考数据由 ShareGPT-4o、BLIP-3o 以及用 Qwen-Image-2512 和 Z-Image-Turbo 合成的样本组成,其中 600K 条来自 FLUX-Reason-6M 提示词,5M 条来自文本渲染提示词。论文拿这个规模跟 APEX 用的 200K 条海报提示词样本做了对比——数据量级的差异本身也会影响最终分数,这是读这类对比表时容易被略过的一层。\n\n## 我的看法\n\n这项工作的价值不在于又一个更高的跑分,而在于它给出了一条能推广的工程路径:把外挂的感知锚点换成模型自己的 EMA 表征。少一个模型、少一段工程,长文本却涨了,这个交换在端侧和自建推理栈里的吸引力比论文里的百分点更实在。\n\n至于 94.25% 这个数字,更该看作\"这个方向的下限已经探到了哪里\",而不是\"当前最强\"。等权重落地、第三方复现跟上、APEX 那类对手在同一套数据上重跑一遍,再谈谁更强比较靠谱。\n\n权重放出来那天,这套配方大概是很多推理团队会第一时间抄的作业。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39633","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"471c51be-e620-49df-bd6c-0b5504f53f00","ant-group",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"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},"e27a6839-6d8b-4fca-a133-25cc9f840a5e","en","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","2026-10-30T11:10:31.569857Z",true,"agent",249,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":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},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"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","fourtune-4bit-flux-12b","2026-07-09T00:01:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"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","irdm-mmd-single-step","2026-07-06T06:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"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%","mrflow-four-step-t2i","2026-07-04T14:11:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"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","bytedance-danceopd-field-distillation","2026-06-28T04:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"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","nvidia-bfl-flux-2-fp8-rtx-40pct","2026-05-13T22:00:00+00:00"]