Research map / Image generation / Text-to-image diffusion
PCPO: Proportionate Credit Policy Optimization for Preference Alignment of Image Generation Models
Jeongjae Lee, Jong Chul Ye
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 proportional timestep weighting to PCPO components (ablation) | PCPO on DDPO · DDPO Aesthetics | training speed | Improved |
| add log rho objective and epsilon-matching before proportional weighting (ablation) | DDPO · DDPO Aesthetics | clipping fraction | Improved |
| apply PCPO to DDPO | DDPO · DDPO Aesthetics (batch 256) | FID: 31.72 -> 27.86 | Improved |
| apply PCPO to DDPO | DDPO · DDPO Aesthetics (batch 512) | FDDINO: -52.831 (p = 0.247) | No clear effect |
| apply PCPO to DanceGRPO | DanceGRPO · FLUX HPS | FID: 46.23 -> 40.38 | Improved |
| apply PCPO to DanceGRPO | DanceGRPO · SD1.4 HPS | LPIPS: 0.4948 -> 0.4894 | No clear effect |
| apply PCPO to DanceGRPO and compare at epoch 120 vs baseline epoch 180 | DanceGRPO · MJHQ-30K FLUX | HPSv2.1: 0.345 -> 0.350 | Improved |
| apply PCPO to DanceGRPO | DanceGRPO · MSCOCO-2017 FLUX | CLIPScore: 0.373 -> 0.369 | Hurt |
| apply PCPO to DDPO BERTScore at low reward targets | DDPO · BERTScore (288/144) | speedup in epochs: 24.2% ... -19.7% | Mixed |
| use timestep subsampling (50% of timesteps) as a heuristic speedup (ablation) | DDPO · DDPO Aesthetics | wall-clock time and image quality: 58% | Mixed |
| emphasize high-noise timesteps with uniform weighting instead of proportional weighting (ablation) | DanceGRPO · DanceGRPO HPS | HPS reward | Hurt |
| use larger batch size 512 vs 256 in DDPO (ablation) | DDPO · DDPO Aesthetics | FID: 31.72 -> 24.09 | Improved |
Closest papers by meaning
- GR-GRPO: Graph-Diffused Credit for Autoregressive Image RL Alignment (ECCV 2026 · similarity 0.88)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.87)
- Curriculum Direct Preference Optimization for Diffusion and Consistency Models (CVPR 2025 · similarity 0.87)
- DrPO: Drifting Preference Optimization for One-Step Generative Models (NeurIPS 2026 · similarity 0.87)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.86)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.86)
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.
