Summary
A paper by Rogerio Guimaraes and Pietro Perona (arXiv:2507.21744) introduces Progressive Seed Pruning (PSP), a new inference-time scaling method for diffusion and flow-matching image generation models. While final image quality is highly sensitive to the initial noise seed, most existing approaches spend extra compute on seed search or resampling under a black-box reward while keeping memory usage constant during inference. The authors show that relaxing this memory constraint opens an underexplored scaling axis: by front-loading exploration, evaluating many seeds early with intermediate denoised estimates, and aggressively pruning weak candidates, a fixed compute budget can be used more effectively. PSP scores intermediate estimates and progressively narrows the candidate set so that only promising trajectories undergo full denoising, keeping the total number of model evaluations constant. Experiments on diffusion and flow-matching backbones show PSP consistently outperforms reward-guided selection baselines including best-of-N, importance sampling, and tree search, achieving higher GenEval scores and better human-evaluated prompt alignment at equal compute.
Paper Overview
- Field: Computer Vision (CV)
- Authors: Rogerio Guimaraes, Pietro Perona
- Published: 2025-07-27
- arXiv: 2507.21744
Summary
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)
- Score intermediate denoised estimates during generation.
- Progressively narrow the candidate set of seeds.
- Only promising trajectories are fully denoised, while the total number of model evaluations stays fixed.
Results
Across diffusion and flow-matching backbones, PSP consistently improves over reward-guided selection baselines at equal compute:
- Higher GenEval scores (automated evaluation)
- Better prompt alignment in human evaluations
- Outperforms best-of-N, importance sampling, and tree search baselines
---
*Auto-collected on 2026-07-27*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178503708