Image generators can improve their own samples without external judges

A training recipe nearly halves a standard image-generation error measure on ImageNet by teaching a model to use previews of its own output.

Industry

Sergey Karpukhin · Aibek Alanov

FusionBrain, AXXX · HSE Unifersity

Research Digest·2 November 2026·2 min read

Illustration: Image generators can improve their own samples without external judges

In one line

A training recipe, Scout-Refine-Correct, lets flow-matching image generators refine their own samples during inference, matching an oracle-based best-of-N baseline on ImageNet without a verifier.

Before you read this

Three ideas this paper builds on

test-time scalingAllocating extra computation during inference to improve output quality.93 papers →self-conditioningThe model uses its own predicted expected embedding to condition subsequent denoising steps.5 papers →diffusion transformerA neural network architecture combining transformer layers with diffusion processes for high-quality image and video generation.63 papers →

Read it as

Image generators usually gain little from extra sampling steps. This work trains one to refine its own output: at each step it looks several steps ahead, forms a tentative image, and feeds that back before the real step; extra rounds repeat this.

A pair of correction losses on the off-path states its rollouts visit make the signal change the sample, not collapse. On ImageNet, FID dropped from the mid-teens to 7.26, matching a privileged many-candidate baseline with a true quality scorer, but no external verifier. Combining both reached 3.81.

FID, Fréchet Inception Distance, measures image quality and diversity; lower is better. Inference-time refinement can be trained directly and adds to sampling-based search.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

flow matching test-time scaling self-conditioning diffusion transformer