Research map / Image generation / Text-to-image diffusion
MPO: Single-Stream Policy Optimization for Efficient Text-to-Image Alignment
Lishuai Gao, Jie Hu, Cong Wei, Yujie Zhong, Yibo Zhao, Zan Gao, Xiaoming Wei
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 |
|---|---|---|---|
| replace group-relative baseline with single-stream Bayesian value tracker | FLUX.1 Dev · GenEval | accuracy: 0.91 -> 0.98 | Improved |
| single-stream MPO instead of DanceGRPO | FLUX.1 Dev · OCR | OCR accuracy: 0.91 -> 0.97 | Improved |
| single-stream MPO instead of DanceGRPO | FLUX.1 Dev · PickScore | PickScore: 23.28 -> 23.89 | Improved |
| single-stream MPO instead of Flow-GRPO | SD3.5-M · GenEval | accuracy: 0.95 -> 0.98 | Improved |
| apply MPO to U-Net Stable Diffusion v1.5 | SD1.5 · GenEval | accuracy: 0.43 -> 0.55 | Improved |
| remove group synchronization (single stream) | FLUX.1 Dev · training cost to GenEval 0.91 | GPU hours: ~235 -> ~9 GPU hours (26x) | Improved |
| single-stream MPO vs group sampling | FLUX.1 Dev · sample efficiency to target GenEval | generated samples: about 5x fewer samples | Improved |
| add global advantage normalization (ablation) | MPO · GenEval | accuracy: 0.93 -> 0.98 | Improved |
| add adaptive forgetting (Q_k from drift proxy) (ablation) | MPO · GenEval | accuracy: 0.94 -> 0.98 | Improved |
| use Bayesian tracker instead of simple EMA baseline (ablation) | MPO · GenEval | accuracy: 0.96 -> 0.98 | Improved |
| add uncertainty-guided prompt curriculum (ablation) | MPO · GenEval | accuracy: 0.96 -> 0.98 | Improved |
| fix initial latent z0 per prompt during training (ablation) | MPO · GenEval | accuracy: 0.93 -> 0.98 (convergence 19.3h -> 9.1h) | Improved |
Closest papers by meaning
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.89)
- GARDO: Reinforcing Diffusion Models without Reward Hacking (NeurIPS 2026 · similarity 0.89)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.88)
- Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models (ICLR 2024 · similarity 0.88)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.87)
- Generalizing Alignment Paradigm of Text-to-Image Generation with Preferences Through f-Divergence Minimization (AAAI 2025 · 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.
