Research map / Image generation / Text-to-image diffusion
Pareto-Guided Optimal Transport for Multi-Reward Alignment
Ying Ba, Tianyu Zhang, Mohan Zhou, Yalong Bai, Wenyi Mo, Guiwei Zhang, Bing Su, Ji-Rong Wen
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 optimal transport mapping to Pareto-bound targets (ablation) | PG-OT · Parti-Prompts | JDR4 (%): 10.78 -> 17.10 | Improved |
| add the online strategy on top of offline strategy (ablation) | PG-OT · Parti-Prompts | JDR2 (%): 38.54 -> 47.98 | Improved |
| use offline Pareto frontier only vs online only (ablation) | PG-OT · Parti-Prompts | JCR4 (%): 4.04 -> 7.23 | Hurt |
| use prompt-specific Pareto-bound targets with OT instead of weighted-sum multi-reward | SD3.5-Turbo + LoRA · Parti-Prompts | JDR2 (%): 26.59 -> 47.98 | Improved |
| optimize a single reward (CLIP only) (ablation) | SD3.5-Turbo + LoRA · CLIP-only optimization | HPS and HP change: HPS -2.78%, HP -4.38% at step 400 | Hurt |
| use prompt-specific bounds via weighted-sum without OT (ablation) | SD3.5-Turbo + LoRA · Parti-Prompts | JDR2 (%): 26.59 -> 29.84 | Mixed |
| use separate-constraint bounds without OT (ablation) | SD3.5-Turbo + LoRA · Parti-Prompts | HPS win rate (%): 82.97 -> 61.21 | Hurt |
| add weak reward models (CLIP, HPS) to joint optimization | multi-reward weighted training · Parti-Prompts | reward hacking | Hurt |
| use PG-OT vs single-reward HP in human study | PG-OT · DiffusionDB | human win rate (%): 76.34 vs SD3.5-Turbo; 61.29 vs HP | Improved |
| add VLM agent to remove weak rewards | PG-OT · Parti-Prompts | JDR2 | Improved |
| vary entropy regularization strength epsilon (ablation) | PG-OT · Appendix E | robustness | No clear effect |
Closest papers by meaning
- Reward Score Matching: Unifying Reward-based Fine-tuning for Flow and Diffusion Models (NeurIPS 2026 · similarity 0.88)
- A Dense Reward View on Aligning Text-to-Image Diffusion with Preference (ICML 2024 · similarity 0.88)
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.87)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.87)
- Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models (ICLR 2024 · similarity 0.87)
- Ctrl-U: Robust Conditional Image Generation via Uncertainty-aware Reward Modeling (ICLR 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.
