Week 28.04.2025 – 04.05.2025

Speculative sampling is a popular technique for accelerating generation in Large Language Models whereby one samples candidate tokens using a fast draft model and accept/reject them based on the target model's distribution. While speculative sampling was previously limited to discrete sequences, we extend it to denoising diffusion models, which are state-of-the-art generative models for image, videos and protein generation. Our experiments demonstrate significant generation speedup on various denoising diffusion models, halving the number of function evaluations, while generating exact samples from the target model. We finally explain how this strategy can be also be used to accelerate simulation of Langevin diffusions.