Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning

Researchers introduced Progressive Seed Pruning, a new inference-time scaling method for diffusion-based image generation.

It scores intermediate denoising states, keeps the most promising seeds, and prunes weaker ones so more candidates can be explored early under a fixed compute budget.

The approach improves prompt alignment and image quality without increasing total model evaluations.

It outperforms several selection baselines in both automated metrics and human evaluation.