Research map / Image generation / Fast diffusion sampling
TDM-R1: Reinforcing Few-Step Diffusion Models with Non-Differentiable Reward
Yihong Luo, Tianyang Hu, Weijian Luo, Jing Tang
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 |
|---|---|---|---|
| apply TDM-R1 RL with GenEval reward | TDM-SD3.5-M (4 NFE) · GenEval overall | score: 0.61 -> 0.92 | Improved |
| apply TDM-R1 with OCR reward | TDM-SD3.5-M (4 NFE) · OCR accuracy | score: 0.55 -> 0.95 | Improved |
| train on GenEval reward | TDM-R1 · DrawBench ImageReward | ImageReward: 0.99 -> 1.11 | Improved |
| train on OCR reward only | TDM-R1 · GenEval | score: 0.61 -> 0.67 | Improved |
| replace TDM-R1 surrogate-reward with direct RL loss (DGPO + TDM distillation) (ablation) | TDM · OCR eval accuracy | OCR accuracy | Hurt |
| use deterministic trajectories instead of stochastic sampling (ablation) | TDM-R1 · OCR eval accuracy | OCR accuracy | Improved |
| use dynamic surrogate reward instead of frozen reward (ablation) | TDM-R1 · OCR eval accuracy | OCR accuracy | Improved |
| use EMA of p_phi as dynamic reference instead of static reference (ablation) | TDM-R1 · OCR eval accuracy | OCR accuracy | Improved |
| RL directly on few-step model rather than distilling a DGPO-finetuned teacher (ablation) | TDM-R1 · OCR eval accuracy | OCR accuracy: 0.91 -> 0.95 | Improved |
| 4-NFE TDM-R1 generator vs 80-NFE surrogate reward diffusion p_phi (ablation) | TDM-R1 · GenEval | score: 0.89 -> 0.92 | Improved |
| apply TDM-R1 to Z-Image 6B with HPSv3 reward | Z-Image-Turbo · HPSv3 | HPSv3: 9.13 -> 9.90 | Improved |
| compare with DGPO/Flow-GRPO on image quality | TDM-R1 · DrawBench Aesthetic | Aesthetic: DGPO 5.31 vs TDM-R1 5.42 | Improved |
Closest papers by meaning
- Reward Fine-Tuning Two-Step Diffusion Models via Learning Differentiable Latent-Space Surrogate Reward (CVPR 2025 · similarity 0.90)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.90)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.89)
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.89)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · similarity 0.89)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.89)
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.
