BWCache: Accelerating Video Diffusion Transformers through Block-Wise Caching

Researchers have developed BWCache, a training-free method that accelerates Diffusion Transformer (DiT) video generation by up to 6× through block-wise feature caching and reuse. The technique exploits computational redundancy in DiT blocks across timesteps while maintaining visual quality, addressing a key bottleneck in real-world AI video generation applications.

• →BWCache achieves up to 6× speedup in DiT-based video generation without requiring model retraining or architectural changes.
• →The method exploits a U-shaped pattern in DiT block feature variations across diffusion timesteps to identify computational redundancy.
• →A similarity indicator triggers feature reuse only when block feature differences fall below a threshold, preserving visual fidelity.
• →DiT blocks are identified as the primary contributors to inference latency in video diffusion models.
• →The training-free approach offers advantages over existing acceleration methods that compromise visual quality through architectural modifications.