Diffusion and flow matching models generate high-quality images but require many iterative sampling steps, making them computationally expensive. Caching-based acceleration offers a training-free way to cut this cost by reusing computations, according to a new paper on arXiv.
PoSTCache, the proposed method, is described as moving "from forecasting to postdictive smoothing" for revisable diffusion caching. This suggests a shift from predicting future states to smoothing past ones, potentially allowing cached results to be revised more effectively.
The abstract notes that caching provides a practical solution, but the full details of the method and its performance are not included in the available excerpt. The paper is available on arXiv.