Keystone judges early diffusion previews to steer image generation
A training-free method scores blurry mid-denoising previews, letting test-time search prune doomed candidates before spending compute on them.
In plain words
Diffusion image generators refine a picture step by step, and each step produces a rough preview of the final image. The Keystone paper says the fate of an image is largely decided early, but existing quality judges misread these blurry previews. Keystone instead turns the prompt into simple questions about objects, counts, colours and positions, which a standard vision-language model can answer reliably.
What was announced
Text-to-image diffusion and flow models produce a clean-image estimate at every denoising step, but test-time search methods have struggled to use these previews. Their verifiers are reward models trained on finished images, so early previews are misjudged, while late judging, as in best-of-N, wastes compute fully denoising candidates that were already doomed, according to the arXiv paper. The authors report that the outcome is largely settled by the seed or the sampler's first steps and is already visible in early previews; the reward models are what fail, reading them near chance on several prompt types.
Keystone, the paper's proposed judge, is training-free: it compiles a prompt into typed questions about objects, counts, colours and positions, and an off-the-shelf multimodal decision model answers them. The authors say this lets the judge read counts and positions from early previews better than reward models read them from finished images. On GenEval with SD3.5-Medium, their adaptive search and pruning reportedly outperformed best-of-8 with every other verifier while using under half its generator compute, and the Keystone score also improved best-of-N, progressive seed pruning and FK steering as a drop-in verifier. Without retuning, both searches stayed above the strongest other verifier's best-of-N on average across four base models and three benchmarks. These are the authors' reported results, not independent replications.
What it does not show
- Evidence is limited to the paper's abstract; the full methodology, experimental detail and any peer review are not covered.
- Reported gains come from the authors' own benchmarks and models, such as GenEval with SD3.5-Medium, and no independent confirmation is available.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.