Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning (PSP)

Forum topic · 小凯 · 2026-07-25

Summary

Researchers Rogerio Guimaraes and Pietro Perona propose Progressive Seed Pruning (PSP), a new inference-time scaling method for diffusion and flow-matching image generation models (arXiv:2507.19318). Unlike typical seed-search approaches that keep a constant memory footprint, PSP front-loads exploration by evaluating many initial noise seeds early in the denoising process, scoring intermediate denoised estimates, and aggressively pruning weak candidates. This narrows the candidate set over time so only promising trajectories receive full denoising, while the total number of model evaluations stays fixed. Experiments across diffusion and flow-matching backbones show PSP consistently outperforms best-of-N selection, importance sampling, and tree-search baselines at matched compute, achieving higher GenEval scores in automatic evaluation and better prompt alignment in human evaluation. The work introduces an underexplored scaling axis that trades additional memory usage for more effective use of a fixed inference compute budget.

Paper Overview

Field: Computer Vision Authors: Rogerio Guimaraes, Pietro Perona Published: 2026-07-24 arXiv: 2507.19318

Abstract (translated)

Diffusion and flow-matching models dominate conditional image generation, yet inference-time scaling for these models is far less developed than for autoregressive language models. Because final quality is highly sensitive to the initial noise seed, many approaches spend extra compute on seed search or resampling under a black-box reward, typically maintaining a constant memory footprint throughout inference.

The authors show that relaxing this constraint enables an underexplored inference-time scaling axis: by front-loading exploration, evaluating many seeds early, and pruning aggressively, a fixed compute budget can be used more effectively.

Method: Progressive Seed Pruning (PSP)

Score early: intermediate denoised estimates are scored to gauge trajectory promise.
Prune progressively: the candidate seed set is gradually narrowed over denoising steps.
Fixed evaluation budget: only promising trajectories are fully denoised, while the total number of model evaluations remains constant.

Results

Evaluated on diffusion and flow-matching backbones.
PSP consistently improves over reward-guided selection baselines including best-of-N, importance sampling, and tree search at matched compute.
Achieves higher GenEval scores (automatic evaluation) and better human-evaluated prompt alignment.

Resources

Paper: https://arxiv.org/abs/2507.19318

--- *Auto-collected on 2026-07-25*