Video diffusion models are powerful but slow, and feature caching has become a popular training-free way to speed them up by reusing intermediate outputs across denoising steps. However, existing adaptive methods decide when to reuse features using heuristic difference measures, which may not be reliable.
The paper argues that heuristics are not enough and introduces a calibrated and corrected caching strategy. This approach is designed to address the limitations of heuristic measures by providing a more principled way to reuse features, all without retraining the model.
The proposed method could make video diffusion acceleration more robust, potentially leading to faster generation while maintaining output quality. The work is available on arXiv and targets a key bottleneck in practical video synthesis.