Research map / Image generation / Super-resolution
GDPO-SR: Group Direct Preference Optimization for One-Step Generative Image Super-Resolution
Qiaosi Yi, Shuai Li, Rongyuan Wu, Lingchen Sun, Zhengqiang Zhang, Lei Zhang
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 GDPO fine-tuning to NAOSD | NAOSD · RealSR | PSNR: 25.25 -> 25.48 | Improved |
| apply GDPO fine-tuning to NAOSD | NAOSD · DIV2K-val | SSIM/LPIPS/FID/DISTS: LPIPS 0.2848 -> 0.2897; FID 25.16 -> 26.44 | Hurt |
| apply GDPO fine-tuning to NAOSD | NAOSD · DIV2K-val | PSNR: 23.83 -> 23.92 | Improved |
| replace Diffusion-DPO with GDPO (ablation) | GDPO-SR · RealSR | PSNR: 25.41 -> 25.48 | Improved |
| replace DanceGRPO with GDPO (ablation) | GDPO-SR · RealSR | PSNR: 25.10 -> 25.48 | Improved |
| replace DanceGRPO with GDPO (ablation) | GDPO-SR · RealSR | AFINE: 16.52 vs 18.72 (lower better) | Hurt |
| add NR metrics to reward (vs FR only) (ablation) | ARF · RealSR | MUSIQ: 67.68 -> 69.42 | Improved |
| add FR metric to reward (vs NR only) (ablation) | ARF · RealSR | LPIPS: 0.2866 -> 0.2675 | Improved |
| add adaptive weighting of FR/NR by smooth vs detail regions (ablation) | ARF · RealSR | CLIPIQA: 0.6331 -> 0.6760 | Improved |
| add GDPO and noise injection to base model | NAOSD · inference cost | time and FLOPs: 0.11s, 2.27T same as OSEDiff | No clear effect |
| use GDPO-SR instead of RL multi-step DP2O-SR | one-step ISR · RealSR | PSNR: 22.39 -> 25.48 | Improved |
Closest papers by meaning
- DP²O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution (NeurIPS 2025 · similarity 0.90)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.86)
- LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution (NeurIPS 2026 · similarity 0.86)
- DrPO: Drifting Preference Optimization for One-Step Generative Models (NeurIPS 2026 · similarity 0.86)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.86)
- Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (CVPR 2024 · 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.
