[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-video-deltanet-h3-hybrid-attention":3,"topics-all":39,"news-related-a2e06c5f-ab65-468f-b936-c654ff75aa38":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":25,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":14,"view_count":38},"a2e06c5f-ab65-468f-b936-c654ff75aa38","14 秒视频 9 秒出片:Video DeltaNet 给 H3 换混合注意力","UC Berkeley 与 Impossible 团队把线性注意力装进视频扩散模型:VDN-H3 在 8 张 B200 上 6.9 秒完成 14.4 秒 768p 视频去噪,约 9 秒返回成片,生成时长跨过播放线。","视频生成一直有个尴尬的算术题:一段十几秒的 768p 视频,扩散模型往往要算上好几分钟,比播放慢一个数量级,直播与交互式生成无从谈起。本周 UC Berkeley、Impossible 与 UT Austin 团队放出的 Video DeltaNet(VDN)把这条裂缝明显推窄:部署在 MiniMax H3 上的 VDN-H3,预热后 8 张 B200 约 9.0 秒返回一段 14.4 秒 768p 成片,去噪循环本身 6.9 秒——生成时长压进播放时长以内(项目页:openvdn.github.io)。\n\n## 注意力是视频生成的头号账单\n\n视频扩散模型去噪时要反复处理长时空 token 序列,softmax 注意力开销随序列长度平方增长。项目页给出扎眼数字:在 H3 这类前沿 omni 模型上,softmax 注意力占总运行时 85% 以上。LLM 那边早有答案——Gated DeltaNet、Kimi Linear 等线性注意力把开销降到线性——但直接照搬到视频模型会翻车:主体一致性、场景布局这类依赖精确交互的全局特征守不住。\n\n## VDN 的拆法:局部 Softmax 加远程线性记忆\n\nVDN 不做二选一,把视频-视频注意力拆成互补两支:局部用滑动窗口 softmax,每 5 个连续潜帧一组、每组只看自己和相邻组,保住细节与短期时序稳定;远程上下文交给双向线性注意力,前向、反向状态各总结窗口前后的帧。另有 4 向边界锚点:每帧都能看到首尾两帧,只增加 3.57% 注意力密度,显著稳住长程一致性。文本与音频相关交互保留 softmax 不动。\n\n真正的新部件是 Video Delta Attention(VDA):把 LLM 里逐 token 的 delta rule 升级成逐帧版本,一帧之内所有空间 token 联立解一个正则化目标,而非各自读写状态——数学上等于对帧级 Gram 矩阵做矩阵逆求解。两个直接好处:继承状态的转移算子范数天然不超过 1,不需要 SANA-WM 那种按帧内 token 数的保守缩放;重复方向的写入互相抑制、独立方向不被稀释。\n\n## 提速账本与三阶段训练\n\n把新分支塞进预训练好的 H3 不能硬训:随机初始化的线性分支会扰动已校准的 softmax 路径。团队三阶段配方:先逐层对齐,再端到端适配(softmax 分支、QKVO 投影与 FFN 全程冻结),最后 LoRA 共适配;softmax 门在前两阶段冻结,防止优化器走\"干脆压制 softmax 分支\"的捷径。\n\n数字层面:单张 B200 上 50 步稠密 H3 去噪 14.4 秒视频要 13.95 分钟,VDN-H3 降到 5.34 分钟。8 卡上 Ulysses 并行把单步延迟从 6.46 秒降到 1.62 秒;再按 profiling 把 5 张卡分给 softmax 分支、3 张给 VDA,进一步降到 1.405 秒,比标准 Ulysses 低 13.3%。叠加 8 步蒸馏(刻意避开激进 4 步)后 DiT 循环 11.23 秒完成——相对稠密单卡基线 74.5 倍、8 卡基线 10.7 倍。摘要同口径:6.70 秒完成 14.3 秒 768p 视频去噪,较稠密基线提速 14.5 倍(arXiv:2609.20744)。\n\n## 所以呢\n\n去噪快于播放时长,流式与直播生成在架构上开始有账可算——论文标题里的 livestream 不是修辞。VDN 与上月底 FastH3 的 4 步蒸馏是两条不同的轴,一个换注意力架构、一个砍步数,原理上可叠加。代码与权重已开放(GitHub 与 Hugging Face 均为 OpenVDN\u002Fvdn-minimax-h3)。留给下一棒的问题很实际:官方\"近无损质量\"的说法,还等第三方复现盖章。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.20744","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,22],{"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":21,"color":14},"f3be854a-50f0-411c-893d-16d0df6def02","h3-series","MiniMax H3 专题：持续追踪 H3 的发布、开源、蒸馏与部署全链路",{"id":23,"name":24,"slug":24,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[26],{"id":27,"lang":28,"title":29,"summary":30,"content":31},"89ada523-ad7b-4645-91f1-025b5925c314","en","Video DeltaNet Makes H3 Video Generation Faster Than Playback","Hybrid attention on MiniMax H3: local softmax plus frame-wise linear memory denoises a 14.4s 768p video in 6.9s on 8 B200s — faster than playback.","Video generation has an awkward piece of arithmetic: a ten-plus-second 768p clip takes diffusion models minutes to render — an order of magnitude slower than playback, which rules out livestream and interactive use. Video DeltaNet (VDN), released this week by a team from UC Berkeley, Impossible, and UT Austin, narrows that gap decisively: VDN-H3, deployed on MiniMax H3, returns a finished 14.4-second 768p video in about 9.0 seconds on 8 B200 GPUs after warm-up, with the denoising loop itself taking 6.9 seconds — generation now runs faster than playback (project page: openvdn.github.io).\n\n## Attention is the dominant bill\n\nVideo diffusion models repeatedly process long spatiotemporal token sequences during denoising, and softmax attention cost grows quadratically with sequence length. The project page puts a stark number on it: on a frontier omni-model like H3, softmax attention accounts for more than 85% of total runtime. LLMs already answered this question — Gated DeltaNet, Kimi Linear and other linear attention families cut the cost to linear — but porting them straight into video models breaks quality: global features like subject identity, scene layout, and temporal dependencies degrade.\n\n## The VDN split: local Softmax plus long-range linear memory\n\nVDN refuses the either\u002For. Video-to-video attention is split into two complementary branches: a sliding-window softmax branch handles local detail — chunks of five consecutive latent frames, each attending to itself and its neighboring chunks — while bidirectional linear attention summarizes distant context, with forward and reverse states covering frames before and after the local window. A 4-way boundary anchor lets every frame see the first and last frames (and vice versa), adding only 3.57% attention density while stabilizing long-range consistency. Interactions involving text or audio keep softmax untouched.\n\nThe genuinely new part is Video Delta Attention (VDA): it upgrades the token-wise delta rule from LLMs into a frame-wise one. All spatial tokens within a frame jointly solve a regularized objective instead of issuing independent reads and writes — mathematically, an (I+A) inverse solve on the frame-level Gram matrix. Two consequences: the inherited-state transition is non-expansive by construction, with no frame-size rescaling needed (unlike SANA-WM), and repeated evidence directions damp each other while independent directions keep full write strength — a hundred patches pointing at one memory direction grow the evidence weight toward, but never past, one.\n\n## The speedup ledger and a three-stage recipe\n\nInserting a randomly initialized linear branch into a pretrained H3 cannot be brute-forced; it perturbs the calibrated softmax path. The team's recipe: layer-wise alignment first, then end-to-end branch adaptation (softmax branch, QKVO projections, and FFN all frozen), then LoRA co-adaptation. The softmax gate stays frozen until the final stage, so the optimizer cannot take the shortcut of simply suppressing the softmax branch.\n\nBy the numbers: on a single B200, 50-NFE dense H3 takes 13.95 minutes to denoise a 14.4-second video; VDN-H3 drops that to 5.34 minutes, with a single block going from 332.5ms to 125.3ms (2.65x). Across 8 GPUs, Ulysses parallelism cuts per-step latency from 6.46s to 1.62s; assigning 5 GPUs to the softmax branch and 3 to VDA (a profiling-driven split) reaches 1.405s, 13.3% below standard Ulysses. With 8-step distillation (deliberately avoiding an aggressive 4-NFE setting), the DiT loop completes in 11.23 seconds — 74.5x versus the dense single-GPU baseline and 10.7x versus the dense 8-GPU baseline. The paper abstract's matching numbers: 6.70 seconds of denoising for a 14.3-second 768p video, a 14.5x speedup over the 50-step dense baseline on the same GPU count (arXiv:2609.20744).\n\n## So what\n\n\"Faster than playback\" is not just a vanity metric: once denoising beats playback duration, streaming and livestream generation becomes economically writable — the livestream in the paper's title is not rhetoric. Note also that VDN and last month's FastH3 (4-step distillation) accelerate on different axes — attention architecture versus step count — and should in principle stack. The full migration path of the DeltaNet family from LLMs into video diffusion, including the three-stage training recipe and SGLang serving, is now open (code and weights: OpenVDN\u002Fvdn-minimax-h3 on GitHub and Hugging Face). The open question for whoever runs next: the official \"near-lossless quality\" claim still awaits third-party replication.","video-deltanet-h3-hybrid-attention","2026-09-18T23:08:51Z","2026-09-18T23:08:59.931382Z","2026-09-18T23:08:59.931393Z",true,"agent",337,[40,49],{"slug":41,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":36,"created_at":47,"modified_at":48},"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":20,"tag_slug":20,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":36,"created_at":55,"modified_at":56},"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},"02c8b500-ec11-44a6-8c58-6e880563dad8","FastH3 开源:4 步蒸馏版 MiniMax H3,B200 单卡最高提速 14 倍","fasth3-4-step-distilled-minimax-h3","2026-08-30T21:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"caed836e-2168-418f-b5c1-bde3ce962e66","Mask Forcing 往蒸馏 rollout 里掺干净 token:修视频生成的模式坍缩,指令遵循最高涨 6.5 分","mask-forcing-video-diffusion-distillation","2026-09-09T23:08:37+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"18d2aa73-7244-4b10-b611-46475e17327e","ForgeWM开源:一步去噪72FPS的可玩世界模型,8张卡复现全流程","forgewm-few-step-playable-world-model","2026-08-24T21:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b5909ee4-586c-494a-9353-4d10dee93227","Reward Lightning:把「打分器」和「蒸馏器」焊进同一根骨干,1-4 步视频生成的同源解法","reward-lightning-video-generation-distillation","2026-07-20T00:15:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"0599b775-ac17-49d2-aebd-a16f531c7168","腾讯混元 MeanFlowNFT：把 RL 接进「平均速度生成器」，Wan 2.1 4 步反超 50 步 LongCat-Video RL","tencent-hunyuan-meanflownft","2026-07-16T12:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"0e44f256-e66e-495c-82e3-aae4dd5e2374","LiveEdit 把扩散视频编辑推到 12.66 FPS：清华让 AR 实时编辑走出 PPT","liveedit-ar-video-editing","2026-07-01T06:15:00+00:00"]