ReHyAt：面向视频扩散 Transformer 的循环混合注意力
ReHyAt: Recurrent Hybrid Attention for Video Diffusion Transformers
把 softmax 注意力的保真度与线性注意力的效率结合成 chunk 级循环形式，并给出从已有双向 softmax 视频扩散模型蒸馏到混合注意力的低成本配方，可直接用于把长视频生成模型扩展到线性复杂度。
原文依据 combines the fidelity of softmax attention with the efficiency of linear attention
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论近期视频扩散模型的进展已转向基于 Transformer 的架构，实现了最优的视频生成效果，但代价是二次方注意力复杂度，严重限制了其向更长序列扩展的能力。我们提出 ReHyAt，一种循环混合注意力机制，它将 softmax 注意力的保真度与线性注意力的效率相结合，支持分块循环重表述与常数内存占用。与同期仅采用线性注意力的 SANA Video 不同，ReHyAt 的混合设计允许从现有基于 softmax 的模型进行高效蒸馏，将训练成本降低两个数量级至 160 GPU 小时，同时在质量上具有竞争力。我们轻量的蒸馏与微调流程提供了一套可应用于未来最优的双向 softmax 模型的配方。在 VBench 与 VBench-2.0 上的实验以及一项人类偏好研究均表明，ReHyAt 在将注意力成本从二次方降至线性复杂度的同时达到了最优的视频质量，为长时长与端侧视频生成释放了实际可扩展性。
英文原文摘要 1145 字符
Recent advances in video diffusion models have shifted towards transformer-based architectures, achieving state-of-the-art video generation but at the cost of quadratic attention complexity, which severely limits scalability for longer sequences. We introduce ReHyAt, a Recurrent Hybrid Attention mechanism that combines the fidelity of softmax attention with the efficiency of linear attention, enabling chunk-wise recurrent reformulation and constant memory usage. Unlike the concurrent linear-only SANA Video, ReHyAt's hybrid design allows efficient distillation from existing softmax-based models, reducing the training cost by two orders of magnitude to 160 GPU hours, while being competitive in the quality. Our light-weight distillation and finetuning pipeline provides a recipe that can be applied to future state-of-the-art bidirectional softmax-based models. Experiments on VBench and VBench-2.0, as well as a human preference study, demonstrate that ReHyAt achieves state-of-the-art video quality while reducing attention cost from quadratic to linear, unlocking practical scalability for long-duration and on-device video generation.