Brief

Researchers have developed a new method to accelerate diffusion models by adapting speculative decoding techniques from large language models. This approach, detailed in a paper on arXiv, introduces a novel scheme that allows for efficient sampling of residual distributions in continuous spaces, a challenge that has previously limited adaptations. The method enables block verification, which provably enhances the acceptance rate of drafts, and formalizes a 'Free Drafter' heuristic that requires no training and offers up to a 6.3% speedup over existing speculative methods. AI