[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-wave-forcing-streaming-video-generation":3,"topics-all":38,"news-related-06b504ab-f9ac-46b2-a1e0-acfbe0fd3996":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},"06b504ab-f9ac-46b2-a1e0-acfbe0fd3996","上交开源WaveForcing:流式生视频134.8帧\u002F秒","上海交大DENG Lab提出Wave Forcing:后续视频块在前面块完成前就去噪,拿更干净的中间结果当上下文,流水线填满后每轮前向产出一个块。8张H200上端到端每秒134.8帧,为单卡7.07倍;百秒视频动态度从36.63升至91.45,代码与权重已开源。","实时生成长视频,卡在两件事上:现有的块式自回归扩散必须等前面的块串行去噪完、再刷新历史缓存,哪怕用一步去噪也得跑两次前向;而 rollout 越长,画质越容易劣化,镜头还会漂移。10 月 31 日,上海交大 DENG Lab 把 [Wave Forcing](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.00859) 挂上 arXiv,三天后登上 Hugging Face Daily Papers 当日榜首(86 赞),给这条流水线做了一次彻底重构。\n\n## 让后面的块提前开工\n\nWave Forcing 的核心是波前调度(wavefront scheduling):后面的视频块不必等前面的块完全去噪,而是用块级因果注意力提前访问前面块更干净的中间状态,把\"更干净到更噪\"的上下文喂给当前块。去掉阻碍流水线化的反向依赖后,系统达到 1F1C——流水线填满之后,每个并发前向轮稳定吐出一个视频块,输出不再是\"等一批\",而是\"匀速流\"。\n\n## 训练和运行时是一起设计的\n\n- **三阶段训练**:混合噪声去噪 → 运动监督 → 因果适应,全程用自生成历史,让模型适应这种\"半成品上下文\"的生成方式。\n- **WaveRT 运行时**:Wave Parallelism 让全模型副本各自处理不同块阶段的任务,逐层 K\u002FV 状态单向发布,通信提前预取、藏进关键 rank 更长的注意力计算里;再叠加优化 kernel 与流水化 VAE 解码。\n\n## 数字层面(论文与仓库自报)\n\n团队在 Wan2.1-T2V-1.3B 基座上蒸馏出 WaveForcing-1.3B:\n\n- **吞吐**:8×H200、832×480 分辨率下,4 步 checkpoint 端到端 134.8 FPS——1593 帧约 11.8 秒解码完,对比匹配条件的单卡 19.06 FPS 基线,提速 7.07 倍;稳态峰值 157.8 FPS。\n- **质量**:5 步 Wave Forcing 与 Rolling Forcing 整体质量相当,但 100 秒长视频的 Dynamic Degree 从 36.63 升到 91.45,Subject Preservation 从 92.75 升到 97.10。\n- **扩展性**:匹配 SageAttention+W8A8、步数与卡数的 DiT-only 测试里,Wave Parallelism 比 Ulysses\u002FRing 两者中更快的一条最高再快 2.53 倍,覆盖 1.3B-14B 六个规模;14B 系统原型跑到 33.5 端到端 FPS(注意:这是随机初始化权重,只测系统扩展性,不代表生成质量)。\n\n## 开源成色\n\nWaveRT 推理代码 Apache 2.0;4 步\u002F5 步两版 checkpoint 放在 Hugging Face,ModelScope 有镜像;训练侧 S1\u002FS2\u002FS3 配方也一并放出(单独的学术许可)。[GitHub 仓库](https:\u002F\u002Fgithub.com\u002FSJTU-DENG-Lab\u002FWaveForcing)目前 91 star。\n\n## 我的看法\n\n这不是\"又一个视频生成模型\",而是一次把调度、训练目标、运行时拧在一起的系统工程——论文标题里的 Speed-of-Light 指向的是延迟,不是质量 SOTA。真正值得盯的是 1F1C 这个性质:生成一旦变成匀速流,直播式内容、可中断可改写的交互视频、甚至世界模型的实时渲染才算有了地基。当然要冷静:以上数字全部团队自报,8×H200 的门槛也不低,第三方复现落地之前,先别急着下\"改朝换代\"的结论。\n\n当\"生成视频\"开始用 FPS 而不是\"像不像\"来打分,这个赛道的下半场,是不是已经提前开始了?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.00859","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"f38fe433-94c8-4201-bc20-6c323faa1a83","en","SJTU Open-Sources Wave Forcing: Streaming Video at 134.8 FPS","SJTU's Wave Forcing streams 134.8 FPS on 8 H200s, 7.07x a single GPU; motion scores on 100-second videos rise from 36.63 to 91.45. Code is open.","Real-time long-video generation is bottlenecked by two things: chunk-wise autoregressive diffusion must serialize denoising and refresh a history cache — two forward passes even with one-step denoising — and longer rollouts degrade quality while the camera drifts. On October 31, DENG Lab at Shanghai Jiao Tong University posted [Wave Forcing](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.00859) to arXiv; three days later it topped Hugging Face Daily Papers as paper of the day (86 upvotes), with a thorough redesign of this pipeline.\n\n## Later chunks start before earlier ones finish\n\nThe core idea is wavefront scheduling: a later video chunk does not wait for earlier chunks to finish denoising. Through chunk-wise causal attention, it accesses the cleaner intermediate states of earlier chunks and feeds that cleaner-to-noisier context into the current chunk. Once the reverse dependencies that block pipelining are removed, the system reaches 1F1C — after pipeline fill, every concurrent forward round steadily produces one output chunk. Output stops arriving in batches and becomes an even stream.\n\n## Training and runtime are co-designed\n\n- **Three-stage training**: mixed-noise denoising, then motion supervision, then causal adaptation, all using self-generated history so the model adapts to generating from half-finished context.\n- **WaveRT runtime**: Wave Parallelism assigns different chunk-stage tasks to full-model replicas, publishes layer-wise K\u002FV states in one direction, and prefetches communication to hide it under the longer attention computation on the critical rank — plus optimized kernels and pipelined VAE decoding.\n\n## The numbers (self-reported by the paper and repo)\n\nThe team distilled WaveForcing-1.3B from the Wan2.1-T2V-1.3B base:\n\n- **Throughput**: on 8×H200 at 832×480, the 4-step checkpoint reaches 134.8 end-to-end FPS — 1,593 frames decoded in about 11.8 seconds, a 7.07× speedup over the matched 19.06 FPS single-GPU baseline; peak steady-state throughput is 157.8 FPS.\n- **Quality**: 5-step Wave Forcing stays comparable to Rolling Forcing overall, while on 100-second videos Dynamic Degree improves from 36.63 to 91.45 and Subject Preservation from 92.75 to 97.10.\n- **Scalability**: in DiT-only tests with matched SageAttention+W8A8 backends, step counts and GPU counts, Wave Parallelism delivers up to 2.53× the throughput of the faster of the Ulysses\u002FRing baselines across six model scales (1.3B-14B); a 14B system prototype reaches 33.5 end-to-end FPS (note: random-initialized weights — this measures system scaling only, not generation quality).\n\n## Open-source substance\n\nWaveRT inference code is Apache 2.0; both the 4-step and 5-step checkpoints are on Hugging Face with a ModelScope mirror; the S1\u002FS2\u002FS3 training recipes are released too (under a separate academic-use license). The [GitHub repo](https:\u002F\u002Fgithub.com\u002FSJTU-DENG-Lab\u002FWaveForcing) currently sits at 91 stars.\n\n## My take\n\nThis is not \"yet another video model\" — it is a systems redesign that fuses scheduling, training objective, and runtime into one loop. Speed-of-Light in the title points at latency, not quality SOTA. The property that really matters is 1F1C: once generation becomes a steady stream, live-style content, interruptible interactive video, and even real-time rendering for world models finally have a foundation. Stay calm, though: every number above is self-reported, eight H200s is a real barrier to entry, and until third-party reproductions land, \"regime change\" conclusions should wait.\n\nWhen generated video starts being scored in FPS rather than \"does it look right\", has the second half of this race already begun?","wave-forcing-streaming-video-generation","2026-11-03T13:05:45Z","2026-11-03T13:05:48.087812Z","2026-11-03T13:05:48.087830Z",true,"agent",90,[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},"eb251970-12c7-4edb-8de5-ef33e60f09fa","BYOD：3.2 GPU 小时把自回归模型改成扩散模型","byod-ar-to-diffusion-conversion-lora-single-gpu","2026-11-02T09:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"144f1f38-77e4-4686-8ffb-2cbed1a5ad12","腾讯开源9B剪辑智能体:token约为Opus 5的3%","tencent-weviscut-long-form-video-editing-agent","2026-11-01T03:05:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"3a336cad-797b-458c-a175-1cd793aec9bf","腾讯混元开源世界模型Hy-Nomas:单卡4090可跑","tencent-hy-nomas-open-world-model","2026-10-28T17:09:08+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"384bb931-49cb-4b1c-b1a8-89495a166699","腾讯ARC Lab:自回归视频的捷径问题,FCP零成本修补","tencent-arc-autoregressive-video-fcp-future-chunk-prediction","2026-10-23T05:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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",{"id":85,"title":86,"news_slug":87,"published_at":88},"05cb84fb-1408-4130-bfe9-9d924461c549","DiffusionART:扩散模型RL微调省12倍GPU时","diffusionart-reward-tilting-gpu-efficiency","2026-10-12T13:15:00+00:00"]