[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-sana-video-2-mixed-linear-attention":3,"news-related-9f443e5f-4ca1-4dfa-b8c7-a5d7ba7aaf6e":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"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","## 先说结论\\n\\n视频生成真正难落地的地方，往往不是画面不够好，而是每次生成都要付出太多注意力计算。7月23日提交 arXiv 的 SANA-Video 2.0，给出的解法不是继续堆参数，而是把注意力机制改成“混合制”：大部分层使用门控线性注意力，保持长序列下的线性扩展；每隔几层再用一次门控 Softmax 注意力，补回纯线性注意力容易丢掉的全局关系。两者按 3:1 交替，再配合 Block Attention Residuals，把前面少数高质量的全局信息传给后续层。\\n\\n论文提供 5B 和 14B 两个规模，目标是单张 GPU 生成最高 720p 视频。实验中，5B 版本在单张 H100 上生成 720p、5 秒视频，完整优化后耗时 13.06 秒；相较匹配的全 Softmax 基线，编译后的 DiT 前向速度提升 3.2 倍，叠加算子融合、缓存和稀疏注意力后再提升 3.58 倍。作者报告的 VBench 分数为 84.30。\\n\\n我的判断是：这篇论文的价值不在“120 倍”这种容易被转发的数字，而在于它把模型结构和推理系统一起设计。线性注意力单独看常常会牺牲表达能力，纯 Softmax 又被长视频的 token 数拖住；SANA-Video 2.0 选择周期性补全全局信息，实际上是在质量和成本之间找可工程化的折中。视频模型下一轮竞争，拼的可能不是谁的演示更惊艳，而是谁能把一次生成的等待时间压进真实工作流。\\n\\n论文仍是预印本，指标也依赖特定硬件和优化栈，不能直接等同于所有场景的加速。但方向很清楚：视频生成要普及，先得把注意力算力省下来。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.21553","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"d4883057-7872-4906-a04c-0eea1a5ff996","en","SANA-Video 2.0: hybrid linear attention, single-GPU video","## Bottom line first Video generation's real deployment blocker isn't the picture quality — it's the attention compute each generation has to pay. SANA-Video 2.0, posted to arXiv on July 23, doesn't keep stacking parameters — it changes the attention mechanism into a \"hybrid\" one: most layers use gated linear attention to keep linear scaling under long sequences, and every few layers insert a gated Softmax attention to recover the global relationships that pure linear attention tends to drop. The two alternate at a 3:1 ratio, paired with Block Attention Residuals, which pass the high-quality global information from the minority of layers on to the layers that follow. The paper provides two scales, 5B and 14B, targeting up to 720p video generation on a single GPU. In experiments, the 5B version generates 720p, 5-second video on a single H100 in 13.06 seconds after full optimization — a 3.2x speedup in compiled DiT forward versus a matched all-Softmax baseline, climbing to 3.58x after operator fusion, caching, and sparse attention stack on top. The authors report a VBench score of 84.30. My take: the value of this paper isn't in the \"120x\" kind of number that's easy to repost — it's in designing the model structure and the inference system together. Linear attention, viewed in isolation, often sacrifices expressiveness; pure Softmax is dragged down by long-video token counts. SANA-Video 2.0's choice to periodically re-supplement global information is essentially finding an engineering-friendly middle ground between quality and cost. The next round of video-model competition may not be about whose demo is more stunning, but about who can push the wait-time of a single generation into a real workflow. The paper is still a preprint, the metrics depend on specific hardware and optimization stacks, and they can't be equated to speedup in every scenario. But the direction is clear: for video generation to become widespread, attention compute has to be saved first.","sana-video-2-mixed-linear-attention","2026-07-24T04:30:00Z","2026-07-24T08:09:31.396842Z","2026-08-19T02:08:40.142862Z",true,"agent",96,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"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":44,"title":45,"news_slug":46,"published_at":47},"8058b748-25c6-4863-917e-46e363773d07","WanToFight 把视频扩散压成 30FPS 实时游戏引擎:多玩家格斗首跑通","wantofight-real-time-game","2026-07-15T16:05:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"b950b487-2b1f-4ece-ad6e-d57cf94f1f84","稀疏注意力新突破：「上下文混合」让长视频生成成本降至近线性","moc-context-mixing-near-linear-video","2026-06-01T01:15:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"94894abf-62aa-41a9-8e3c-e999ff274d60","Sparse Forcing：稀疏注意力让视频生成质量速度双提升","meta-ucsb-sparse-forcing-pbsa-video-1-27x","2026-05-07T08:10:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"0bc5be19-abbb-4898-8e1d-86cb127fbb8a","长上下文调优重塑视频生成：单镜头模型学会「讲连续的故事」","lct-long-context-tuning-video-scene-consistency","2026-05-03T10:05:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f4c705fd-47c9-481a-807f-8001820070f8","InfinityEdit:三注意力轻量适配器,把视频编辑推进无界流时代","infinityedit-infinite-video-editing-adapter","2026-08-25T13:00:00+00:00"]