ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE
Researchers have developed ReCache, a novel method for optimizing caching schedules in diffusion models to improve image and video generation efficiency. This technique uses policy gradients to learn a recomputation schedule that maximizes generation quality within a specified computational budget, without requiring labeled data. ReCache demonstrates significant improvements over existing methods, reducing LPIPS and boosting VBench scores on benchmarks like FLUX and Wan 2.1. AI
IMPACT Improves efficiency of diffusion models, potentially lowering compute costs for generative AI applications.