[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-diffusionart-reward-tilting-gpu-efficiency":3,"topics-all":35,"news-related-05cb84fb-1408-4130-bfe9-9d924461c549":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"05cb84fb-1408-4130-bfe9-9d924461c549","DiffusionART:扩散模型RL微调省12倍GPU时","arXiv 2610.12572 提出 DiffusionART:把提示词采样奖励聚合为共享改进方向,免 EMA 调参与可微奖励。它在混元视频 13B 上比 DiffusionNFT 少用 12.02 倍 GPU 时,训练奖励高 28.8%,10 月 12 日登顶 HF Daily Papers。","给视频扩散模型做强化学习微调,是一笔让工程团队头疼的算力账:奖励上去了,GPU 时也烧没了。10 月 8 日提交到 arXiv 的论文 [DiffusionART](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.12572) 给出了一套新解法:论文报告在 HunyuanVideo-13B 上,达到 DiffusionNFT 后期平均奖励水平所需的 GPU 时减少 12.02 倍,同时平均训练奖励还高出 28.8%。10 月 12 日,这篇论文以 151 赞排在 Hugging Face Daily Papers 当日列表首位。\n\n## 背景:两条路线各有各的坑\n\n用黑盒奖励在线微调扩散模型,现有方法分两条路线。反向过程方法以 Flow-GRPO 为代表,论文指出它受制于信用分配不准,效率上不去;前向过程方法以 DiffusionNFT 为代表,效率高,但离开精心调节的 EMA(指数移动平均)离策略训练就不稳定。\n\n论文的核心洞察是:NFT 式目标里藏着一股\"前向-反向一致性力\",正负样本联合训练只能部分抵消它;而 EMA 之所以能稳住训练,靠的是\"恢复\"机制,而不是离策略采样本身。\n\n## DiffusionART 的做法\n\nDiffusionART 全称 Diffusion Reinforcement with Aggregation-First Reward Tilting。它把每个提示词采样得到的所有奖励先聚合成一个共享的\"奖励改进方向\",再作用于反向过程的状态上——一致性力在构造层面就被移除,不再需要靠训练去对抗。在此基础上,它加入一次带信任域掩码的第二步更新,自适应修复训练漂移,整套流程不再需要调节 EMA。\n\n工程上有三点值得注意:一是不需要可微分的奖励模型,黑盒奖励即可;二是每个提示词所需采样数更少仍能保持效果;三是支持成对奖励模型,论文给出了 DiffusionART-Pairwise 变体。论文称在图像与视频生成任务上,它无需按任务调参即可高效稳定地训练。代码已随论文放出([GitHub 仓库](https:\u002F\u002Fgithub.com\u002Fvanzll\u002FDiffusionART))。\n\n## 成绩单:12 倍 GPU 时的差距\n\n在 HunyuanVideo-13B 上,以视觉质量与运动质量等权重的奖励设置对比:对比 DiffusionNFT,DiffusionART 达到其后期平均奖励水平所用 GPU 时少 12.02 倍,平均训练奖励高 28.8%;对比 Flow-GRPO-Fast,GPU 时少 11.72 倍,平均训练奖励高 53.0%。\n\n对做视频生成的团队来说,这个数字的含义很直接:同样的算力预算,迭代轮次可以多出一个数量级。当奖励模型是黑盒、算力是硬约束时,训练方法本身的效率差异,会被放大成产品迭代速度的差异。\n\nHugging Face Daily Papers 页面将其列于 KlingAI Research 名下。社区对它的关注点也很明确:不依赖 EMA 调参、不需要可微奖励,意味着这套流程对工程落地的门槛大幅降低。\n\n所以值得盯的不是某一个\"12 倍\",而是这条曲线:RL 微调扩散模型的单位成本正在快速下降——当微调便宜到可以随手做,差异化就从\"谁有大集群\"转向\"谁有好奖励\"。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.12572","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",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},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"97831198-7ebd-4731-aa21-33d24715affd","en","DiffusionART: 12x Fewer GPU Hours for Diffusion RL","DiffusionART removes a hidden consistency force in diffusion RL: 12.02x fewer GPU hours than DiffusionNFT on HunyuanVideo-13B, +28.8% reward.","Reinforcement-learning fine-tuning for video diffusion models has always been a painful GPU-hours bill: rewards go up, and the compute budget burns down with them. [DiffusionART](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.12572), a paper submitted to arXiv on October 8, offers a new recipe: the authors report that on HunyuanVideo-13B it reaches the late-stage mean reward level of DiffusionNFT with 12.02x fewer GPU hours, while scoring 28.8% higher mean training rewards. On October 12 the paper topped the Hugging Face Daily Papers listing with 151 upvotes.\n\n## Background: Two Routes, Two Traps\n\nOnline RL fine-tuning of diffusion models with black-box rewards has split into two camps. Reverse-process methods, exemplified by Flow-GRPO, are limited in efficiency by poor credit assignment, as the paper points out. Forward-process methods, exemplified by DiffusionNFT, are efficient but unstable without carefully tuned exponential-moving-average (EMA) off-policy training.\n\nThe paper's central insight: NFT-style objectives hide a forward-reverse consistency force that joint positive-negative training only partially cancels, and EMA stabilizes training through restoration rather than off-policy rollouts.\n\n## The DiffusionART Recipe\n\nDiffusionART stands for Diffusion Reinforcement with Aggregation-First Reward Tilting. It first aggregates all sampled rewards for each prompt into a shared reward-improving direction, then applies that direction to reverse-process states — the consistency force is removed by construction instead of being fought during training. On top of this, it adds a trust-region-masked second update for drift-adaptive restoration, so the whole pipeline no longer needs EMA tuning.\n\nThree engineering details stand out. First, no differentiable reward model is required — black-box rewards suffice. Second, it stays effective with fewer samples per prompt. Third, it supports pairwise reward models via the DiffusionART-Pairwise variant. Across image and video generation tasks, the authors report efficient and stable training without per-task tuning. Code is available in the [GitHub repository](https:\u002F\u002Fgithub.com\u002Fvanzll\u002FDiffusionART).\n\n## The Scorecard: A 12x Gap in GPU Hours\n\nOn HunyuanVideo-13B, with equally weighted visual and motion quality rewards: versus DiffusionNFT, DiffusionART reaches its late-stage mean reward level with 12.02x fewer GPU hours and 28.8% higher mean training rewards; versus Flow-GRPO-Fast, it needs 11.72x fewer GPU hours and scores 53.0% higher.\n\nFor video-generation teams the implication is blunt: the same compute budget now buys roughly an order of magnitude more iterations. When reward models are black boxes and compute is the hard constraint, differences in training-method efficiency compound into differences in product iteration speed.\n\nThe Hugging Face Daily Papers page lists the paper under KlingAI Research. What the community is watching is equally clear: no EMA tuning and no differentiable reward means the engineering barrier to adoption drops sharply.\n\nSo the thing to watch is not any single \"12x\" — it is the curve: the unit cost of RL fine-tuning for diffusion models is falling fast. When fine-tuning becomes cheap enough to do casually, differentiation shifts from \"who owns a big cluster\" to \"who owns a good reward\".","diffusionart-reward-tilting-gpu-efficiency","2026-10-12T13:15:00Z","2026-10-12T13:13:16.020725Z","2026-10-12T13:13:16.020739Z",true,"agent",121,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"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":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"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":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"989509cd-0977-4ddf-9443-8ef49485befd","ViRDM:三网络视频蒸馏,砍到只剩一个生成器","virdm-teacher-free-video-post-training","2026-09-26T19:06:45+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"9f443e5f-4ca1-4dfa-b8c7-a5d7ba7aaf6e","SANA-Video 2.0：用混合线性注意力把视频生成推到单卡可用","sana-video-2-mixed-linear-attention","2026-07-24T04:30:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"79ed2e02-2fe4-43ca-a9b2-847740969424","HDR 把视频模型的多步推理硬拉出新手感:层级隐变量让经典规划任务成功率从 34% 跳到 60%","hdr-video-multi-step-planning","2026-07-18T12:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"13d7d706-62ea-4adc-aa72-6e064f2fa5d5","弱模型给强模型当导航:阿里扩散训练法把FID压到1.15","swift-diffusion-target-extrapolation-weak-to-strong","2026-10-21T05:01:38+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"4349edac-997e-4980-95a0-58054a28a9dd","视频生成投机去噪:1.7 步顶 4 步,提速 1.5 倍","speculative-forcing-video-denoising","2026-10-16T19:11:48+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"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"]