Research map / Reinforcement learning / RL from feedback
Test-Time Reinforcement Learning for Flow Matching
Jili Chen, Changqin Huang, Qionghao Huang, Yaxin Tu, Zhonglong Zheng, Xiaodi Huang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add KL constraint to the reference latent (ablation) | Flow-TTRL (SD3.5-M) · GenEval | Overall score: 0.71 -> 0.87 | Improved |
| add KL constraint to the reference latent (ablation) | Flow-TTRL (Flux.1 Dev) · GenEval | Overall score: 0.69 -> 0.83 | Improved |
| add PRDP warm-up stage 1 (ablation) | Flow-TTRL (SD3.5-M) · GenEval | Overall score: 0.81 -> 0.87 | Improved |
| add PRDP warm-up stage 1 (ablation) | Flow-TTRL (Flux.1 Dev) · GenEval | Overall score: 0.76 -> 0.83 | Improved |
| add GRPO refinement stage 2 (ablation) | Flow-TTRL (SD3.5-M) · GenEval | Overall score: 0.72 -> 0.87 | Improved |
| add GRPO refinement stage 2 (ablation) | Flow-TTRL (Flux.1 Dev) · GenEval | Overall score: 0.70 -> 0.83 | Improved |
| add GRPO refinement stage 2 (ablation) | Flow-TTRL (SD3.5-M) · Pick-a-Pic | PickScore: 22.07 -> 22.24 | Improved |
| add noise reweighting (ablation) | Flow-TTRL (SD3.5-M) · GenEval | Overall score: 0.85 -> 0.87 | Improved |
| add noise reweighting (ablation) | Flow-TTRL (Flux.1 Dev) · GenEval | Overall score: 0.82 -> 0.83 | Improved |
| use dynamic noise reweighting instead of static noise level alpha (ablation) | Flow-TTRL · GenEval | Overall score | Improved |
| use very small or very large static noise alpha (ablation) | Flow-TTRL · GenEval | Overall score | Hurt |
| increase group size G beyond 8 (ablation) | Flow-TTRL · T2I-CompBench Complex | Complex score | No clear effect |
Closest papers by meaning
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.92)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.90)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.88)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.88)
- GARDO: Reinforcing Diffusion Models without Reward Hacking (NeurIPS 2026 · similarity 0.87)
- The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL. (NeurIPS 2026 · similarity 0.87)
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.
