CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration
Training-free feature reuse accelerates diffusion transformer inference by replacing full denoising computation at selected steps with cached approximations.
arXiv cs.LG2 min read