MEND: RL for flow models via proximal velocity matching beats Flow-GRPO in 100 vs ~4k updates

UTEXAS · hf · 2026-10-07

UT Austin researchers introduce MEND, a reinforcement learning method for reward post-training of flow models built on proximal velocity matching.

Method

Caps rewards within each prompt group so well-scoring samples receive no update.
Below the cap, proposes moves along the reward gradient and accepts one only when its capped reward gain exceeds a quadratic displacement price.
Regresses onto resulting velocity targets with no KL term, frozen reference model, or advantage weights.

Results

In 100 updates, MEND outperforms Flow-GRPO (4k updates) on five of six evaluators at the same distance to base-model images.
Under equal budget, it surpasses ReFL and DiffusionNFT at every evaluated update across four rewards, reaching PickScore 24.03 vs 23.92 and 23.43.
A 300-update three-reward run beats the five-reward DiffusionNFT model on all three rewards it trains on.

The method is general and applies to any flow backbone with a differentiable reward.

More from Multimodal