Research map / Image generation / Super-resolution
Learning Coupled Dictionaries from Unpaired Data for 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 DictSR, which learns coupled HR and LR dictionaries via an autoencoder to associate unpaired HR and LR images in a compact code space for ×4 image super-resolution, instead of synthesizing pseudo LR images.
Problem
Training learning-based image super-resolution on unpaired real-world HR and LR images, where existing methods rely on synthesizing pseudo LR images whose realness and diversity are limited by the large image space.
Approach
Associates unpaired HR and LR images in a compact proxy space by discretizing encoded features into coupled HR content, LR content, and hierarchical degradation dictionaries (shared spatial codes via a downscaled HR bridge), trained end-to-end with an autoencoder; the SR network is a shallow 18-layer encoder-decoder.
Key results
- On AIM-RWSR (x4), DictSR reaches LPIPS 0.259, PSNR 22.46, SSIM 0.629, best among unpaired methods (DeFlow: 0.349/22.25/0.620).
- On NTIRE-RWSR (x4), DictSR achieves LPIPS 0.204 vs DeFlow 0.218, with PSNR 25.90 vs 25.87 and SSIM 0.711 vs 0.710.
- DictSR uses an 18-layer network with 129G MACs, about half the MACs of 350-layer ESRGAN-based unpaired methods (291G).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DictSR instead of Real-ESRGAN | AIM-RWSR ×4 · AIM-RWSR | LPIPS: 0.238 -> 0.259 | Hurt |
| use DictSR instead of BSRGAN | AIM-RWSR ×4 · AIM-RWSR | PSNR: 22.47 -> 22.46 | Hurt |
| use DictSR instead of DeFlow | NTIRE-RWSR ×4 · NTIRE-RWSR | LPIPS: 0.218 -> 0.204 | Helped |
| remove degradation dictionary Ddeg (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 22.40 | Hurt |
| reduce dictionary entries to 2048 (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 22.41 | Hurt |
| increase dictionary entries to 8192 (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 22.44 | Hurt |
| increase latent dimension to 16 (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 22.25 | Hurt |
| increase latent dimension to 64 (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 21.99 | Hurt |
| reduce LR/HR dictionary patch sizes to 1×1/4×4 (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 22.41 | Hurt |
| increase LR/HR dictionary patch sizes to 4×4/16×16 (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 21.43 | Hurt |
| remove shared spatial codes coupling HR and LR dictionaries (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 8.28 | Hurt |
| deepen HR encoder to 20 layers (ablation) | DictSR · AIM-RWSR | PSNR: 22.46 -> 22.46 | No clear effect |
Surprising findings
- Removing the coupling operation between HR and LR dictionaries drops performance to near-zero quality (PSNR 8.28), showing coupling is essential.
- Deepening the encoders and decoders gives no consistent gains, suggesting the dictionary design matters more than network capacity.
Limitations
- Relies on comparisons with GAN-based and paired baselines on only two synthetic unpaired benchmarks.
- Performance on NTIRE-RWSR over DeFlow is marginal in PSNR and SSIM.
Experimental setup
Datasets: AIM-RWSR NTIRE-RWSR DIV2K
Benchmarks: AIM-RWSR NTIRE-RWSR
Metrics: PSNR SSIM LPIPS
Baselines: RCAN ZSSR IKC DAN BSRGAN Real-ESRGAN CycleGAN CinCGAN DASR DeFlow FSSR PCR-ESRGAN
Builds on: VQ-VAE VQGAN coupled dictionaries for sparse coding SR
Closest papers by meaning
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.88)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.88)
- GSDD: Generative Space Dataset Distillation for Image Super-resolution (AAAI 2024 · similarity 0.87)
- Lightweight Image Super-Resolution via Flexible Meta Pruning (ICML 2024 · similarity 0.87)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.86)
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (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.