Diffusion Model Caching

Diffusion model caching exploits the fact that some model computations change little between denoising steps or across video frames, so they can be reused instead of recomputed. Methods cache intermediate features or attention key-value states, forecast skipped features from past steps, or selectively refresh them using signals such as sensitivity, motion, uncertainty, and error estimates. Approaches range from training-free schedules to learned predictors and adaptive policies, and are evaluated on image and video generation benchmarks such as COCO and VBench as well as diffusion-language and long-context tasks. The central challenge is balancing speed and memory savings against errors that accumulate through generation, especially when content, model stages, or sampling conditions change.