Research map / Image generation / Super-resolution
DP²O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DP²O-SR, a DPO-style post-training framework for generative real-world super-resolution that builds preference pairs from multiple noise-seed outputs ranked by a hybrid full-reference/no-reference IQA reward, with hierarchical weighting (HPO).
Problem
Generative Real-ISR models built on T2I diffusion/flow backbones produce outputs of varying perceptual quality across noise seeds, and existing Diff-DPO uses only a single best-vs-worst pair from different models, leaving this stochastic perceptual range underexploited.
Approach
Treat the stochasticity of T2I sampling as a source of preference supervision: sample M outputs per LR input from one model, rank them with a hybrid FR+NR IQA reward, build N² pairs from top-N vs bottom-N, and weight pairs adaptively by intra-group reward gap and inter-group reward spread (HPO).
Key results
- On RealSR, DP²O-SR (SD2) reaches MANIQA 0.7031 and CLIPIQA+ 0.7852, the highest among compared methods.
- On the Syn-Test set, DP²O-SR (FLUX) improves MANIQA from 0.6857 to 0.7199 and QALIGN from 4.4266 to 4.7060 over C-FLUX.
- In a user study with 10 participants, DP²O-enhanced models were preferred over their baselines in 67.5% (C-SD2) and 69.5% (C-FLUX) of pairwise comparisons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DP²O-SR preference optimization | C-SD2 · RealSR | MANIQA: 0.6629 -> 0.7031 | Helped |
| Apply DP²O-SR preference optimization | C-SD2 · RealSR | CLIPIQA+: 0.7295 -> 0.7852 | Helped |
| Apply DP²O-SR preference optimization | C-FLUX · RealSR | MANIQA: 0.6632 -> 0.6918 | Helped |
| Apply DP²O-SR preference optimization | C-FLUX · RealSR | LPIPS: 0.4004 -> 0.4024 | Hurt |
| Apply DP²O-SR preference optimization | C-SD2 · Syn-Test | MANIQA: 0.6684 -> 0.7165 | Helped |
| Add intra-group pair weighting (ablation) | HPO · RealSR | Perceptual Reward: 0.645 -> 0.648 | Helped |
| Add inter-group weighting (ablation) | HPO · RealSR | Perceptual Reward: 0.645 -> 0.649 | Helped |
| Add both intra-group and inter-group weighting (ablation) | HPO · RealSR | Perceptual Reward: 0.645 -> 0.651 | Helped |
| Use lower selection ratio such as 1/16 rather than higher ratios (ablation) | C-FLUX preference curation · RealSR | Perceptual reward | Mixed |
| Use selection ratio N/M = 1/4 (ablation) | C-SD2 preference curation · RealSR | Perceptual reward | Helped |
| Increase rollout count M (ablation) | DP²O-SR preference curation · RealSR | Perceptual reward | Helped |
Surprising findings
- PSNR and SSIM decrease after DP²O-SR training (e.g., RealSR PSNR for SD2 drops from 23.61 to 22.49), consistent with the perception-distortion tradeoff.
- Optimal selection ratio depends on model capacity: smaller C-SD2 prefers broad coverage (N/M=1/4), while larger C-FLUX prefers stronger contrast (N/M=1/16).
Limitations
- IQA-based reward lacks interpretability and does not fully capture subjective perceptual quality.
- Training pipeline is fully offline; no iterative or online preference optimization.
- PSNR/SSIM fidelity declines after preference training.
Experimental setup
Datasets: RealSR ResShift first-order degradation pipeline curated 30,100-image internet dataset
Benchmarks: RealSR Syn-Test (100 held-out images)
Metrics: LPIPS TOPIQ-FR AFINE-FR MANIQA MUSIQ CLIPIQA+ TOPIQ-NR AFINE-NR QALIGN VQ-R1
Baselines: StableSR DiffBIRv2 SeeSR CCSR AddSR OSEDiff
Builds on: Diff-DPO ControlNet DPO
Closest papers by meaning
- LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution (NeurIPS 2026 · similarity 0.92)
- Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment (NeurIPS 2025 · similarity 0.90)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.88)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.88)
- Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.88)
- SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution (CVPR 2024 · 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.