Research map / Image generation / Super-resolution
RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution
Yushuai Song, Weize Quan, Weining Wang, Jiahui Sun, Jing Liu, Meng Li, Pengbin Yu, Zhentao Chen, Wei Shen, Lunxi Yuan, Dong-ming Yan
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 |
|---|---|---|---|
| fine-tune Qwen3-VL 8B with GRPO on LR-conditioned rank rewards | Qwen3-VL 8B evaluator · RefSR-18K in-domain | Agreement (%): 31.8 -> 85.0 | Improved |
| use no-think fast inference instead of full reasoning (ablation) | RefReward-SR · RefSR-18K in/out-of-domain | Agreement (%): 85.0 -> 84.9 (in), 80.2 -> 79.8 (out) | No clear effect |
| add global-local crop scoring (ablation) | RefReward-SR · RefSR-18K in/out-of-domain | Agreement (%) | Improved |
| scale preference training data from 1K/2K groups to full (ablation) | RefReward-SR · RefSR-18K | Agreement (%) | Improved |
| GRPO fine-tune SR model with RefReward-SR + LPIPS + DEQA rewards | C-FLUX · RealSR | Q-Align: 3.4943 -> 4.1596 | Improved |
| GRPO with RefReward-SR rewards versus DPO fine-tuning | C-FLUX · RealSR/DIV2K-Val | NR metrics: MUSIQ 72.42 -> 71.70 (RealSR); Q-Align 3.9757 -> 4.1596 | Mixed |
| add RefReward-SR reward (ablation) | C-FLUX GRPO · RealSR | PSNR: 23.89 -> 24.68 | Improved |
| add RefReward-SR reward (ablation) | C-FLUX GRPO · RealSR | NR metrics (MUSIQ): 73.48 -> 71.70 | Hurt |
| add LPIPS reward (ablation) | C-FLUX GRPO · RealSR | PSNR / LPIPS: PSNR 21.98 -> 24.68; LPIPS 0.4026 -> 0.3202 | Improved |
| add DEQA reward (ablation) | C-FLUX GRPO · RealSR | MUSIQ: 69.02 -> 71.70 | Improved |
Closest papers by meaning
- LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution (NeurIPS 2026 · similarity 0.93)
- Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment (NeurIPS 2025 · similarity 0.91)
- DP²O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution (NeurIPS 2025 · similarity 0.90)
- Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.89)
- SRSR: Enhancing Semantic Accuracy in Real-World Image Super-Resolution with Spatially Re-Focused Text-Conditioning (NeurIPS 2025 · similarity 0.88)
- RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (AAAI 2025 · similarity 0.88)
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.
