Research map / Image generation / Super-resolution
Unsupervised Diffusion-Based Degradation Modeling for Real-World 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 UDDM, an unsupervised diffusion-based degradation modeling framework that synthesizes realistic LR-HR training pairs from unpaired real LR and HR images, using a diffusion degradation module (DDM) bridged by extreme downsampling plus a physics-based dynamic degradation module (P-DDM), to train SR models.
Problem
Real-world super-resolution suffers a domain gap because SR models trained on synthetic degradations do not match real-world LR degradation distributions, and paired real LR-HR data is costly and misaligned.
Approach
Use extreme (ultra-low-resolution) downsampling as a bridge between unpaired real LR and HR images so a diffusion model learns the real LR degradation distribution (diffusing real LR to ULR, reverse-diffusing HR-derived ULR to LR), then fuse it with a physics-based, content-adaptive degradation branch to produce pseudo pairs for training SR.
Key results
- On RealSR with StableSR, the proposed data pipeline reaches PSNR 26.820 and LPIPS 0.2514 versus 25.679 and 0.3680 for Syn-Real.
- On DRealSR with SwinIR, PSNR 29.247 and LPIPS 0.2709 versus 28.301 and 0.3801 for Syn-Real, and 26.944 and 0.3219 for Real-ESRGAN degradation.
- Fusion weight w=0.5 (UDDM) balances fidelity and perceptual quality; DRealSR PSNR 28.03 vs 27.43 for DDM alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UDDM fusion with w = 0.5 instead of DDM alone (w = 0.0) (ablation) | UDDM · RealSR and DRealSR | PSNR: RealSR: 26.10 -> 26.63; DRealSR: 27.43 -> 28.03 | Helped |
| use UDDM fusion with w = 0.5 instead of P-DDM alone (w = 1.0) (ablation) | UDDM · RealSR and DRealSR | PSNR: RealSR: 26.44 -> 26.63; DRealSR: 27.97 -> 28.03 | Helped |
| use UDDM fusion with w = 0.5 instead of DDM alone (w = 0.0) (ablation) | UDDM · RealSR and DRealSR | SSIM: RealSR: 0.7626 -> 0.7821; DRealSR: 0.7341 -> 0.7936 | Helped |
| use UDDM fusion with w = 0.5 instead of P-DDM alone (w = 1.0) (ablation) | UDDM · RealSR and DRealSR | SSIM: RealSR: 0.7716 -> 0.7821; DRealSR: 0.7540 -> 0.7936 | Helped |
| use UDDM fusion with w = 0.5 instead of DDM alone (w = 0.0) (ablation) | UDDM · RealSR and DRealSR | LPIPS: RealSR: 0.3190 -> 0.2439; DRealSR: 0.3595 -> 0.2654 | Helped |
| use UDDM fusion with w = 0.5 instead of P-DDM alone (w = 1.0) (ablation) | UDDM · RealSR and DRealSR | LPIPS: RealSR: 0.2503 -> 0.2439; DRealSR: 0.3080 -> 0.2654 | Mixed |
| use UDDM fusion with w = 0.5 instead of DDM alone (w = 0.0) (ablation) | UDDM · RealSR and DRealSR | CLIP-IQA: RealSR: 0.6127 -> 0.6234; DRealSR: 0.6340 -> 0.6357 | Mixed |
| use UDDM fusion with w = 0.5 instead of P-DDM alone (w = 1.0) (ablation) | UDDM · RealSR and DRealSR | CLIP-IQA: RealSR: 0.5847 -> 0.6234; DRealSR: 0.5893 -> 0.6357 | Helped |
| use UDDM fusion with w = 0.5 instead of DDM alone (w = 0.0) (ablation) | UDDM · RealSR and DRealSR | MUSIQ: RealSR: 65.81 -> 65.88; DRealSR: 58.98 -> 58.51 | Mixed |
| use UDDM fusion with w = 0.5 instead of P-DDM alone (w = 1.0) (ablation) | UDDM · RealSR and DRealSR | MUSIQ: RealSR: 64.05 -> 65.88; DRealSR: 56.77 -> 58.51 | Helped |
Surprising findings
- Real-ESRGAN's own degradation model attains slightly higher SSIM and FID than UDDM on some settings, which the authors attribute to trivial differences.
Limitations
- High time and computational cost due to diffusion sampling (authors' stated limitation)
- Evaluation limited to two real-world benchmarks, RealSR and DRealSR
- Fusion weight w must be tuned, trading fidelity against perceptual quality
Experimental setup
Datasets: DIV2K Flickr2K OutdoorSceneTraining RealSR DRealSR
Benchmarks: RealSR DRealSR
Metrics: PSNR SSIM LPIPS FID CLIP-IQA MUSIQ
Baselines: Real-ESRGAN degradation Syn-Real SwinIR Real-ESRGAN StableSR
Builds on: StableSR Real-ESRGAN DDPM Syn-Real
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.92)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.91)
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.91)
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · similarity 0.91)
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.91)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.90)
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.