Progressive Seed Pruning: Inference-Time Scaling for Diffusion Models (arXiv 2507.20484)

Forum topic · 小凯 · 2026-07-26

Summary

Researchers Rogerio Guimaraes and Pietro Perona (Caltech) propose Progressive Seed Pruning (PSP), a new inference-time scaling method for diffusion and flow-matching image generation models. Because final image quality is highly sensitive to the initial noise seed, prior approaches spend extra compute on seed search or resampling under black-box rewards while keeping memory usage constant. PSP relaxes that constraint: it front-loads exploration by evaluating many seeds early, scores intermediate denoised estimates, and progressively prunes weaker candidates so that only promising trajectories are fully denoised—all within a fixed compute budget. Experiments across diffusion and flow-matching backbones show PSP consistently outperforms reward-guided selection, achieving higher GenEval scores (automated) and better human-evaluated prompt alignment than best-of-N, importance sampling, and tree-search baselines at equal compute. Project page: https://www.vision.caltech.edu/psp. Code: https://github.com/rogerioagjr/psp. Paper: arXiv:2507.20484.

Paper Overview

Field: Computer Vision (CV)
Authors: Rogerio Guimaraes, Pietro Perona
Published: 2026-07-25
arXiv: 2507.20484

Abstract (translated from the original)

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, but typically maintain 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) scores intermediate denoised estimates and progressively narrows the candidate set so that only promising trajectories are fully denoised, while keeping the total number of model evaluations unchanged.

Results

Across diffusion and flow-matching backbones, PSP consistently outperforms reward-guided selection: at the same compute budget, it achieves higher GenEval scores (automated evaluation) and better human-evaluated prompt alignment than best-of-N, importance sampling, and tree-search baselines.