Research map / Image generation / Super-resolution
Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning
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 LWay, a training framework that pre-trains an LR reconstruction network on real paired data and then self-supervisedly fine-tunes off-the-shelf SR models on unseen real test LR images using a DWT-weighted high-frequency loss.
Problem
Real-world image super-resolution generalizes poorly from synthetic degradations to real test images, and real HR/LR pairs in the target domain are costly or unavailable.
Approach
Pre-train a degradation encoder and HR-to-LR reconstructor, freeze them, then fine-tune any pretrained SR model on test LR images by reconstructing the LR input from the SR output and a DWT-weighted high-frequency loss, requiring no paired target HR data and no architecture changes.
Key results
- On RealSR Nikon, Real-ESRGAN+ with LWay improves PSNR by 1.77 dB and SSIM by 0.0388 over the baseline.
- FeMaSR with LWay gains +2.44 dB PSNR and -0.0796 LPIPS on RealSR Canon (Table 1).
- A user study with 24 researchers found LWay scored more than 2 points above the second-best method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LWay self-supervised fine-tuning to BSRGAN | BSRGAN · RealSR Canon, Nikon, and Sony; DRealSR Olympus and Panasonic | PSNR: +0.04; +0.52; +0.76; +0.24; +1.07 | Helped |
| Add LWay self-supervised fine-tuning to BSRGAN | BSRGAN · RealSR Canon, Nikon, and Sony; DRealSR Olympus and Panasonic | LPIPS: 0.1988 -> 0.1572; 0.2302 -> 0.1750; 0.2685 -> 0.2469; 0.3275 -> 0.2926; 0.2738 -> 0.2624 | Helped |
| Fine-tune each image individually rather than not fine-tuning | BSRGAN + LWay · RealSR Nikon | LPIPS: 0.3136 -> 0.2351 | Helped |
| Fine-tune the entire 10-image test set rather than not fine-tuning | BSRGAN + LWay · RealSR Nikon | LPIPS: 0.3136 -> 0.2536 | Helped |
| Add 40 same-sensor images to fine-tuning (50 images total) (ablation) | BSRGAN + LWay · RealSR Nikon | LPIPS: 0.2536 -> 0.2571 | Mixed |
| Increase LR reconstruction training data from 2K to 6k synthetic images (ablation) | LR reconstruction network · RealSR Nikon | LPIPS: 0.1836 -> 0.1816 | Helped |
| Increase LR reconstruction training data from 0.6K to 4K real-world images (ablation) | LR reconstruction network · RealSR Nikon | LPIPS: 0.2003 -> 0.1722 | Helped |
| Increase LR reconstruction training data from 4K to 6K real-world images (ablation) | LR reconstruction network · RealSR Nikon | LPIPS: 0.1722 -> 0.1800 | Hurt |
| Vary degradation embedding dimension (128, 256, 512, 1024, 2048, or 4096) (ablation) | BSRGAN + LWay · RealSR Nikon | LPIPS: 0.1804; 0.1776; 0.1722; 0.1736; 0.1629; 0.1669 | Mixed |
| Fine-tune the baseline with real paired data (ablation) | BSRGAN · RealSR Nikon | LPIPS: 0.2302 -> 0.2268 | No clear effect |
| Add high-frequency weighted loss (ablation) | BSRGAN + LWay · RealSR Nikon | LPIPS: 0.1858 -> 0.1722 | Helped |
Surprising findings
- Supervised fine-tuning on real paired data from one sensor gives only marginal gains and can yield over-smoothed outputs on another sensor's data.
- Fine-tuning on a single image gives the best results, but collective fine-tuning on 10 images is more efficient and still gives significant gains.
Limitations
- Single-image fine-tuning, the best-performing setting, is costlier than collective fine-tuning.
- Gains on some sensor/model combinations are small (e.g., Sony and olympus LPIPS changes near zero for some models).
- LR reconstruction network relies on in-house real paired data that is not released.
Experimental setup
Datasets: RealSR DRealSR in-house real paired images (6,000)
Benchmarks: RealSR DRealSR
Metrics: LPIPS DISTIS NLPD PSNR SSIM MAD
Baselines: Real-ESRGAN+ BSRGAN SwinIR-GAN FeMaSR StableSR ZSSR DASR DiffBIR DARSR LDM CAL GAN
Builds on: Real-ESRGAN BSRGAN ZSSR StableSR
Closest papers by meaning
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.91)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.89)
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · similarity 0.89)
- Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution (CVPR 2024 · similarity 0.88)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.88)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.87)
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.