Show HN: HiCache++ — training-free diffusion speedup, lossless ~2× further than HiCache (DMD/Prony basis)
A new method called HiCache++ has been developed to significantly speed up diffusion models, such as those used in Stable Diffusion, without requiring additional training. This technique improves upon existing methods by using Dynamic Mode Decomposition (DMD) and Prony basis, which are better suited for forecasting the trajectory of diffusion features compared to the polynomial basis used in earlier approaches. HiCache++ demonstrates improved performance and maintains lossless quality at wider skip intervals, making it a drop-in replacement for current caching mechanisms. AI
IMPACT This new method could lead to faster generation times for diffusion models, making them more accessible and efficient for users.