AI Paper Atlas

Research map / Image generation / Text-to-image diffusion

TEMPFLOW-GRPO: WHEN TIMING MATTERS FOR GRPO IN FLOW MODELS

Xiaoxuan He, Siming Fu, Yuke Zhao, Wanli Li, Jian Yang, Dacheng Yin, Fengyun Rao, Bo Zhang

ICLR 2026 · Image generation › Text-to-image diffusion · approach: Preference optimization for diffusion

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace Flow-GRPO with TempFlow-GRPOSD3.5-M · GenEvaloverall score: 0.88 -> 0.97 at 3800 stepsImproved
TempFlow-GRPO convergence speedSD3.5-M · GenEvalsteps to 0.95: about 2,000 vs 5,600 stepsImproved
replace Flow-GRPO with TempFlow-GRPOSD3.5-M · PickScorePickScore: +1.7% vs Flow-GRPO, +1.0% vs Flow-GRPO (Prompt)Improved
replace Flow-GRPO with TempFlow-GRPOFLUX.1-dev · HPSv3 rewardsteps to match: 300 -> 80 stepsImproved
add noise-aware policy reweighting (ablation)Flow-GRPO · GenEvalscore at 1200 steps: 0.82 -> 0.92Improved
add trajectory branching (ablation)Flow-GRPO · GenEvalscore: approx +5%Improved
add seed group strategy (ablation)TempFlow-GRPO · GenEvalscore: approx +2%Improved
add trajectory branching, noise reweighting, then seed group (ablation)Flow-GRPO (Prompt) · PickScorePickScoreImproved
more initial noise seeds with fewer branches (6x4 vs 2x12) (ablation)TempFlow-GRPO · group size 24training curveMixed
replace Flow-GRPO (Prompt) with TempFlow-GRPO under multi-rewardFLUX.1-dev · multi-reward PickScore promptsHPSv3: 10.41 -> 12.23Improved
replace DanceGRPO with TempFlow-GRPOHPSv2 reward · HPDv2HPSv2 score: 37.2 -> 38.5Improved
use TempFlow-GRPO at 1024 resolutionSD3.5-M 1024 · PickScoresteps to match: 450 -> about 100 stepsImproved

Later papers that compared with this method

1 later accepted paper reports TempFlow-GRPO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy