Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning

Researchers Rogério Guimarães and Pietro Perona at Caltech introduce Progressive Seed Pruning, or PSP, a training-free, gradient-free method for scaling diffusion and flow-matching inference under a fixed denoising budget. Instead of fully denoising a constant number of samples, PSP starts with a larger pool of noise seeds, scores intermediate clean-image estimates using the black-box ImageReward model, and progressively retains only the most promising trajectories. Its default schedule uses an effective compute multiplier of 4, begins with 8 seeds, prunes to 4 at 25 percent progress, and then to 2 at 50 percent. On GenEval prompts, PSP reaches GenEval scores of 0.574 for Stable Diffusion v1.5, 0.645 for SDXL, and 0.747 for Stable Diffusion 3.5, outperforming Best-of-N, FK-Steering, and DSearch at matched compute; human prompt-alignment evaluation gives SDXL a score of 0.713. PSP remains effective with deterministic samplers, which is important for flow-matching models, and its intermediate rewards can be cached to search pruning schedules offline. On an H200, SDXL PSP at multiplier 4 takes 13.74 seconds, close to 12.48 seconds for Best-of-4 and substantially below 23.90 seconds for Best-of-8. The authors also extend PSP to prompt selection using 24 rewrites generated by OpenAI’s ChatGPT 5.2, suggesting the method can optimize any discrete candidate pool rather than only random seeds.