Diffusion and flow models generate a sample over iterative steps, and each step requires a full pass through a large network. That pass dominates the overall generation cost. Training-free caching has emerged as a way to accelerate this process by reusing activations from recent steps, avoiding redundant computation.
SoftCache, described in a new arXiv preprint, proposes adaptive step size caching specifically for diffusion transformers. The method adaptively reuses activations from recent steps, aiming to cut the number of expensive full passes while preserving the quality of the generated sample. Because it is training-free, it can be applied to existing models without additional fine-tuning.
The abstract is brief and does not include experimental details or quantitative results, so the reported effectiveness and exact mechanism for adapting step size are not yet available from this source. The paper is listed as a new announcement on arXiv, and further details would be needed to assess its practical impact.