[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-faab4a6c-9cb0-4f2a-a5bf-1f122306008b":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"faab4a6c-9cb0-4f2a-a5bf-1f122306008b","阿里 Wan-Streamer v0.2 把分辨率从 192p 推到 640p:Thinker-Performer 拓扑保住 200ms 实时延迟","阿里 Wan Team 在 arXiv 上线 Wan-Streamer v0.2(arXiv:2607.04443),是一次\"延迟不变、分辨率翻几倍\"的工程升级。v0.1 已经把全双工音频-视频交互塞进单 Transformer 的统一因果时间线,代价是输出只有 192×336,够做视频通话近景,放到中景里人物姿态、桌面物件、周围环境全部糊成一片。\n\nv0.2 的目标:分辨率从 192×336 抬到 640×368(像素量约 3.5 倍),帧率仍 25 FPS,模型侧信号到信号延迟保持 ~200ms,含 350ms 双向网络预算的总远程交互延迟维持 ~550ms。这意味着升级不能动那条对延迟敏感的因果路径,新增算力只能向非延迟关键环节分流。\n\n解法是 Thinker-Performer 部署拓扑的重新切分。Thinker 继续驻留在单卡,负责流式感知、短语言\u002F状态 Transformer、KV-cache 构建、最后一拍解码;Performer 改为 Ulysses 式上下文并行的多卡组,专门承担长序列潜空间去噪:每个 rank 维护按 Ulysses 分片的本地 KV-cache,高分辨率潜视频序列在 rank 间做 all-to-all\u002Fgather,短音频潜不分片。Thinker 只把 performer 可消费的 KV slice 广播过去,语言状态本身不需要再跨卡同步,远端延迟守在 ~550ms 区间。\n\n视觉层面,v0.2 让近景通话更清晰,首次支持\"场景内中景数字人\":坐姿、眼神、手部动作、桌面物品在实时对话里保持可读,数字人不再被锁在画脸框里。这是把全双工交互从\"对话\"推进到\"在场景里对话\"的一步,对客服坐席、虚拟主播、陪伴机器人都直接影响成片观感。\n\n更大的视角下,这条路径说明实时音视频生成不再是\"推理速度优化\",而是进入\"流式因果 + 部署拓扑协同设计\"阶段——next-unit 流式建模、context-parallel performer、低延迟 thinker 守护,三者已形成完整工程模板,Grok、GPT-4o 这类实时语音+视觉系统迟早要走同样的拓扑分层。","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2607.04443v3","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation","2026-07-10T16:15:00Z","2026-07-10T16:15:33.442393Z","2026-07-10T16:15:33.442402Z",true,"agent",2]