Jiajun Fan

Keystone: Judging Previews at Key Steps for Test‑Time Search in Diffusion Models

  • 1UIUC
  • 2NYU
  • 3CMU

At a glance

The paper in four findings.

Test-time search for text-to-image models finishes every candidate before it judges it. Keystone judges candidates at the beginning, from the model’s own blurry previews, and spends compute only where the outcome is still open.

  • Training-free.a released judge; no step labels, no reward model trained on noise
  • Reads previews.the model’s own clean-image estimate, at no extra network evaluations
  • Plugs in.any diffusion or flow sampler, any search that keeps candidates

The problem

Search finishes every image before it judges it, even the ones doomed from the start.

Text-to-image models still miscount, drop and misplace objects. The standard remedy, best-of-N, draws several complete images and keeps the one a verifier prefers. The verifiers are reward models trained on finished images, so the generator pays in full for every candidate the search will throw away.

Best-of-8 on “a photo of four bananas”every candidate runs all 40 steps before any is judged

Reward models reading the step-8 previewcandidate pairs ordered correctly

SD3.5-Medium, development prompts. Right: fraction of candidate pairs with different official verdicts that each judge orders correctly from the step-8 preview.

The image was never the problem. The judge was.

The idea

The layout is in place long before the picture is: ask the preview the prompt’s own questions.

Every denoising step yields the model’s own estimate of the finished image. Decoded, it is a blurry preview, but how many objects there are, which ones and where is already visible.

Two candidates for “a photo of four bananas”the Keystone score of each preview above it

Keystone changes the judge instead of training one.

How it works

Compile the prompt, read at the key steps, search where it matters.

Three parts, and none of them is trained.

  • no training
  • no step labels
  • no gradient through the judge
  • no change to the sampler
  1. 1Compile

    The prompt becomes typed questions: presence and count, colour, position.

  2. 2Read

    At a key step the decision model answers every question from the preview; the score is the smallest answer.

  3. 3Search

    Adaptive search reads one candidate at step 16 and finishes it at 0.85 or above, else draws five more and finishes the best. Pruning keeps four of eight at step 8 and two at step 16, then compares the two finished images.

The previews’ answers decide which candidates receive further compute; nothing is differentiated through the judge and the sampler is unchanged.

The evidence Finding 01

Decided in the first steps, readable by step 8.

Change only the noise of one step and see how often the official verdict changes; then ask how well each judge orders candidates from the preview at each step.

Change the noise of one stephow often the official verdict changes

Order candidates from the previewcandidate pairs ordered correctly, by the step that is read

SD3.5-Medium, development prompts. Left: 40 sibling groups per step. Right: counting and two-object prompts pooled, 40-step Euler sampler.

Search should branch at the start and read shortly after.

The evidence Finding 02

A step-8 preview, read better than a finished image.

On identical candidates, the Keystone score reads counts and positions from step-8 previews better than reward models read them from finished images.

Countingcandidate pairs ordered correctly

Positioncandidate pairs ordered correctly

SD3.5-Medium, development prompts (the paper’s Table 1). The Keystone score reads the step-8 preview; the other judges read the finished image.

Counting

a photo of three handbags

Step-8 previews and their Keystone scores; four kept

Three handbags in a row: black, pink and brownKeystone’s pruning ends onCorrect
Two handbags, orange and purpleImageReward’s pruning ends onWrong

Position

a photo of a laptop left of a cow

Step-8 previews and their Keystone scores; four kept

A laptop on the left and a cow on the rightKeystone’s pruning ends onCorrect
A cow standing in a field behind a laptopImageReward’s pruning ends onWrong

Colour attribution

a photo of a black broccoli and a yellow cake

Step-8 previews and their Keystone scores; four kept

A slice of yellow cake beside dark broccoliKeystone’s pruning ends onCorrect
Green broccoli on top of a yellow cakeImageReward’s pruning ends onWrong

Examples: official test prompts, chosen by the paper’s fixed rule (App. D.16.2).

Reading the preview well is what lets search stop early.

The evidence Finding 03

More accurate than every other verifier’s best-of-8, at under half its compute.

On GenEval with SD3.5-Medium, both searches lie above best-of-N with every judge at the same cost.

GenEval overall against computenetwork evaluations per image, log scale

Counting and positionGenEval, one sample against Keystone pruning

Official test prompts, four images per prompt. Best-of-8 costs 320 network evaluations per image.

Pruning reaches 0.82 at 144 network evaluations; every other verifier’s best-of-8 costs 320 and stays at or below 0.78.

The evidence Finding 04

A drop-in judge, across models and benchmarks.

Swap the judge inside existing methods, or move to other models and benchmarks without retuning.

Best-of-N

Better than every other verifier at every N from 2 to 8.

Progressive seed pruning

+0.071

with the Keystone score in place of ImageReward

FK steering

0.709 → 0.754+0.045

with the Keystone score in place of ImageReward

SD3.5-Medium

An open book with a quill penA man in a brown hat, seen from behindA green and an orange jack-o’-lantern

FLUX.1-dev

A steam locomotive behind potted plants on a windowsillA hot dog on an orange suitcaseA red and black city bus on a street

SDXL

A pink and green motorcycle in front of a pink wallA yellow kitchen with a keyboard on the counterA brown suitcase on the deck of a boat

Sana-0.6B

Airplanes flying in a blue skyRows of colourful carsA brown backpack and a blue cow

BenchmarksGenEvalT2I-CompBenchGenAI-Bench

Drop-in: GenEval overall on SD3.5-Medium, same candidates, only the judge changes. Other models: a language model writes each prompt’s slots once; each benchmark’s own metric scores the images.

Averaged over models, both searches beat the strongest other verifier’s best-of-N on all three benchmarks.

Why it matters

Test-time search no longer has to finish an image to judge it.

Compute where the outcome is still open.

Adaptive search draws a second batch for 42% of images; the rest finish after one candidate.

No training on noise.

A released judge reads the sampler’s own previews, so any diffusion or flow sampler with a clean-image estimate can be searched at its key steps.

A signal at every step.

The same judgment of previews could supply the per-step signal that reinforcement learning of flow models still lacks.

The film

The story, as a short film.

3 min 57 s, with narration. The voice is synthesized.

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}
}