BWCache: Accelerating Video Diffusion Transformers through Block-Wise Caching

简介

扩散 Transformer（DiT）是目前最好的视频生成方法，但逐步去噪的过程很慢，限制了实际应用。我们的分析发现，DiT 模块是推理延迟的主要来源；而且在去噪过程中，各模块的特征变化呈 U 形：开始和结束阶段变化大，中间阶段相邻步的特征高度相似，存在大量冗余计算。

BWCache 是一种无需训练的加速方法：在去噪步之间动态地缓存并复用各 DiT 模块的特征，并用一个相似度指标判断何时复用，只有相邻步的模块特征差异低于阈值时才复用，在减少冗余计算的同时保持画质。在 Open-Sora、Open-Sora-Plan、Latte、Wan2.1、HunyuanVideo 等模型上的实验表明，BWCache 最高加速 2.6 倍，画质与原模型相当。

方法

Abstract（英文摘要）

Recent advancements in Diffusion Transformers (DiTs) have established them as the state-of-the-art method for video generation. However, their inherently sequential denoising process results in inevitable latency, limiting real-world applicability. Existing acceleration methods either compromise visual quality due to architectural modifications or fail to reuse intermediate features at proper granularity. Our analysis reveals that DiT blocks are the primary contributors to inference latency. Across diffusion timesteps, the feature variations of DiT blocks exhibit a U-shaped pattern with high similarity during intermediate timesteps, which suggests substantial computational redundancy. In this paper, we propose Block-Wise Caching (BWCache), a training-free method to accelerate DiT-based video generation. BWCache dynamically caches and reuses features from DiT blocks across diffusion timesteps. Furthermore, we introduce a similarity indicator that triggers feature reuse only when the differences between block features at adjacent timesteps fall below a threshold, thereby minimizing redundant computations while maintaining visual fidelity. Extensive experiments on several video diffusion models demonstrate that BWCache achieves up to 2.6× speedup with comparable visual quality.

引用

@inproceedings{cui2026bwcache,
  title     = {BWCache: Accelerating Video Diffusion Transformers through Block-Wise Caching},
  author    = {Cui, Hanshuai and Tang, Zhiqing and Xu, Zhifei and Yao, Zhi and Zeng, Wenyi and Jia, Weijia},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}

其他项目