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

Forum topic · 小凯 · 2026-07-26

Summary

This forum post introduces the paper 'Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning' (arXiv 2507.20484) by Rogerio Guimaraes and Pietro Perona (Caltech). Diffusion and flow-matching models dominate conditional image generation, but their inference-time scaling lags behind autoregressive language models. Since final image quality is highly sensitive to the initial noise seed, existing methods spend extra compute on seed search or reward-guided resampling while keeping memory constant. The authors relax this constraint and propose Progressive Seed Pruning (PSP): front-load exploration by evaluating many seeds early, score intermediate denoised estimates, and aggressively prune the candidate set so only promising trajectories are fully denoised, all under a fixed total compute budget. Across diffusion and flow-matching backbones, PSP consistently outperforms reward-guided selection, achieving higher GenEval scores 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 Overview

Field: Computer Vision Authors: Rogerio Guimaraes, Pietro Perona arXiv: 2507.20484

Abstract (translated from the original forum post)

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.

Progressive Seed Pruning (PSP)

PSP scores intermediate denoised estimates and progressively narrows the candidate set, so that only promising trajectories are fully denoised.
The total number of model evaluations remains unchanged compared to baseline methods.

Results

Evaluated on diffusion and flow-matching backbones.
PSP consistently outperforms reward-guided selection, achieving higher GenEval scores (automated evaluation) and better prompt alignment in human evaluation than best-of-N, importance sampling, and tree-search baselines at equal compute.