Diffusion models deliver strong image and video generation, but their inference cost is high because every sampling step requires a full denoiser evaluation. Feature caching is an attractive, deployment-friendly way to cut these costs, but existing approaches can degrade quality. A new arXiv preprint proposes a training-free alternative that forecasts denoiser outputs along a Bezier curve, stabilized with Tikhonov regularization.

The method, described in arXiv:2610.16448, targets Diffusion Transformers specifically. Instead of retraining or modifying the network, it predicts how features evolve across steps and reuses cached values where appropriate. The authors argue that a Bezier representation gives a flexible yet smooth path for forecasting, while Tikhonov stabilization keeps the prediction well-conditioned, reducing drift and artifacts.

Because the approach is training-free, it can be dropped into existing pipelines without additional optimization. The abstract does not report quantitative results, so claims about speedup or quality preservation are not available from this source alone.