Keystone: Judging Previews at Key Steps for Test-Time Search in Diffusion Models
In brief
What the paper does.
An image model decides what it will draw in its first few steps, yet test-time search finishes every candidate before judging it, because the usual judges are reward models built for finished images. Keystone judges candidates at those key steps instead, from the model’s own blurry previews: an off-the-shelf decision model answers questions compiled from the prompt, and each image is scored by its weakest answer. Search can then stop finishing images that were doomed from the start. On GenEval with Stable Diffusion 3.5 Medium, it is more accurate than best-of-8 with every other verifier at under half its compute. Training-free, and a drop-in judge for existing search methods; without retuning, the gain holds on average across four base models and three benchmarks.
Cite
Cite the paper.
@article{fan2026keystone,
title={Keystone: Judging Previews at Key Steps for Test-Time Search in Diffusion Models},
author={Fan, Jiajun and Cheng, Chaoran and Wang, Ziwen and Yang, Changyi and Yang, Bangji and Zheng, Qinglong and Liu, Ge},
journal={arXiv preprint arXiv:2611.25530},
year={2026},
url={https://arxiv.org/abs/2611.25530}
}