Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps

This paper delves into the inference-time scaling of diffusion models, investigating how generation performance can be improved with increased computation beyond simply increasing denoising steps.

The authors formulate the challenge as a search problem over sampling noises, using verifiers to evaluate noise candidates and algorithms to find better noise candidates.

The study reveals that increasing inference-time computation leads to substantial improvements in the quality of samples generated by diffusion models, and it discusses the optimal use of verifiers and algorithms in different application scenarios.