training-free
[CVPR 2024] DeepCache: Accelerating Diffusion Models for Free
Drop-in TaylorSeer/HiCache basis upgrade — training-free diffusion acceleration via a Dynamic Mode Decomposition (Prony) exponential feature-forecast basis. Not the SGLang KV-cache HiCache.
T-GATE: Temporally Gating Attention to Accelerate Diffusion Model for Free!