PoSTCache improves cache accuracy for faster image generation
PoSTCache: From Forecasting to Postdictive Smoothing for Revisable Diffusion Caching
Computer Vision and Pattern Recognition
Summary
Creating images with diffusion models takes a lot of computing time because the process uses many steps. The authors show a way to speed this up by not only predicting intermediate images but also revising their guesses after seeing better information later. This lets the model fix earlier shortcuts, making the final images look better without extra training. They also designed a smart method to choose which steps to do fully to get the best improvements. Their experiments show this approach works well for generating images and videos.
What this means in practice
- •For graphics software developers: Enable faster generation of high-quality images by revising intermediate cached features during diffusion sampling without retraining models.
- •For video content creators: Speed up video generation by improving how intermediate steps are corrected and integrated in diffusion-based video models.
Authors
Liwei Lan, Hui Chen, Jungong Han, Guiguang Ding
Abstract
Diffusion and flow matching models achieve impressive visual generation quality but remain computationally expensive due to iterative sampling. Caching-based acceleration provides a practical training-free solution by reusing or predicting intermediate features. However, existing methods largely enforce a forward-only information flow: approximations produced for skipped model evaluations are immediately committed to the sampling trajectory and cannot benefit from feedback generated by later full-computation steps. We propose \textbf{PoSTCache} (\emph{Postdictive Smoothing Trajectory Cache}), a training-free and plug-and-play framework that reformulates diffusion caching as revisable trajectory inference. PoSTCache first uses an existing caching strategy to construct forward proposals for skipped steps. At a later full-computation step, the discrepancy between the refreshed prediction and its forward proposal yields a correction signal, which PoSTCache propagates backward to refine and reintegrate the preceding predictions. Viewing full-computation steps as sources of corrective feedback also changes how they should be selected. We therefore introduce the Posterior Variance Reduction (PVR) strategy, which estimates cross-timestep dependencies from offline denoising trajectories and selects anchors according to their expected reduction in surrogate trajectory uncertainty rather than local approximation error alone. Experiments on representative image and video generation models show that PoSTCache consistently improves the quality--efficiency trade-off over representative caching methods.