Iterative Partial Refinement Boosts Diffusion Model Inference Scaling

A new method called Iterative Partial Refinement (IPR) improves inference-time scaling for diffusion models without requiring external verifiers. Proposed in a paper on arXiv (2605.19317), IPR re-noises and regenerates subsets of an already-generated sample, conditioning on remaining regions to revise earlier decisions under richer context. This addresses a gap in existing scaling methods that rely on external reward models, which limit scalability. IPR is tailored for sequential diffusion models with region-wise, mixed-noise conditioning, an underexplored area. The approach enables self-correction during inference, potentially enhancing reasoning capabilities in generative AI.