# Discriminator-Guided RL Sharply Improves Flow-Matching Diffusion Model Quality Without Human Preference Labels

_Research · published 2026-06-22_

Researchers from Meta AI and collaborators have proposed Discriminator-Guided RL (DRL), a method to correct a structural flaw they identify in flow- and score-matching diffusion models. The core argument is that matching losses—which minimize regression error on velocity or score fields—are a poor proxy for the visual and semantic quality that actually matters at inference time. Rather than rely on expensive human preference annotations, DRL trains a discriminator to separate real data from base-model samples in a pretrained representation space, then uses its logit as a reward signal in KL-regularized reinforcement learning.

The approach yields substantial gains across multiple backbone architectures (SiT, JiT, REPA, and RAE). On SiT, guidance-free FID drops from 9.38 to 2.62, and semantic-space FD on DINOv3 falls from 88.2 to 19.3. Notably, DRL also improves human-preference scores without ever training on human preference data, and when followed by preference-based post-training it achieves a better Pareto frontier between alignment and image fidelity—reducing artifacts like oversaturation in the process. The paper, submitted to arXiv on 17 June 2026, runs to 84 pages including appendices.

## Sources
- [arxiv.org](https://arxiv.org/abs/2606.19162)
---
Canonical: https://genbuzz.news/posts/discriminator-guided-rl-sharply-improves-flow-matching-diffusion-model-quality-without-human-preference-labels