Wave Forcing：面向光速流式视频生成的协同设计框架
原标题：Paper page - Wave Forcing: Towards Speed-of-Light Streaming Video Generation
AI 摘要
上海交通大学 Deng Lab 提出 Wave Forcing（WF），一种面向流式长视频生成的训练与执行框架，通过分块因果注意力、错峰去噪与缓存刷新调度，以及三阶段训练流程，实现连续长视频输出。在 Wan2.1-T2V-1.3B 上，5 步 WF 质量与 Rolling Forcing 相当，100 秒视频的 Dynamic Degree 从 36.63 提升至 91.45；在 8 块 H200 上 4 步 WF 达到 134.8 端到端 FPS，为单 GPU 吞吐的 7.07 倍。其 Wave Parallelism 在 DiT 测试中最高达 Ulysses/Ring 基线吞吐的 2.53 倍，代码、训练配方与检查点已开源。
正文节选
Wave Forcing: Towards Speed-of-Light Streaming Video Generation Abstract Wave Forcing enables efficient continuous long-video generation with consistent appearance and natural motion through co-designed generation, training, and execution, achieving high throughput and quality. Real-time long-video generation requires continuous output with consistent appearance and natural motion. Existing chunk-wise autoregressive methods face challenges in both generation efficiency and long-video quality: se