Research map / Image generation / Super-resolution
AND: Adversarial Neural Degradation for Learning Blind Image Super-Resolution
Fangzhou Luo, Xiaolin Wu, Yanhui Guo
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 |
|---|---|---|---|
| add adversarial perturbation, neural degradation and identity init to classical bicubic SR training (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | PSNR of ANDNet: 26.53 -> 28.47 | Improved |
| add adversarial perturbation, neural degradation and identity init to classical bicubic SR training (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | LPIPS of ANDGAN: 0.4245 -> 0.2326 | Improved |
| add neural degradation (with identity init) to adversarial noise training (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | PSNR of ANDNet: 26.60 -> 28.47 | Improved |
| replace additive adversarial noise with adversarial neural degradation (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | LPIPS of ANDGAN: 0.4194 -> 0.2326 | Improved |
| add adversarial perturbation to neural degradation with identity init (vs synthetic augmentation) (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | PSNR of ANDNet: 27.31 -> 28.47 | Improved |
| add adversarial perturbation to neural degradation with identity init (vs synthetic augmentation) (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | LPIPS of ANDGAN: 0.3089 -> 0.2326 | Improved |
| add identity initialization of the degradation network instead of random Xavier init (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | PSNR of ANDNet: 11.84 -> 28.47 | Improved |
| adversarial noise training instead of classical bicubic training (adversarial noise only) (ablation) | AND (ANDNet/ANDGAN) · RealSR (x4) | PSNR of ANDNet: 26.53 -> 26.60 | No clear effect |
| ANDNet vs DCLS, DAN, DASR baselines on four real-world SR benchmarks | ANDNet · RealSR, DRealSR, SupER, ImgPairs | PSNR: 27.83 -> 28.47 (RealSR vs DCLS) | Improved |
| ANDGAN vs perceptual baselines (SwinIR-Real, DASR, Real-ESRGAN) | ANDGAN · RealSR, DRealSR, SupER, ImgPairs | LPIPS: 0.2515 -> 0.2326 (RealSR vs SwinIR-Real) | Improved |
| ANDNet LPIPS vs best baseline on DRealSR and SupER | ANDNet · DRealSR / SupER | LPIPS: 0.2739 vs 0.3381 (DRealSR) | Hurt |
| adversarial degradation training cost vs plain ESRGAN training | ESRGAN · training cost per step | TFLOPs: 8.92 -> 57.85 TFLOPs (6.49x) | Hurt |
Closest papers by meaning
- Learning Correction Filter via Degradation-Adaptive Regression for Blind Single Image Super-Resolution (ICCV 2023 · similarity 0.92)
- Unsupervised Degradation Representation Learning for Blind Super-Resolution (CVPR 2021 · similarity 0.91)
- Learning a Single Convolutional Super-Resolution Network for Multiple Degradations (CVPR 2018 · similarity 0.90)
- End-to-End Alternating Optimization for Real-World Blind Super Resolution (IJCV 2023 · similarity 0.90)
- DDR: Exploiting Deep Degradation Response as Flexible Image Descriptor (NeurIPS 2024 · similarity 0.88)
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · 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.
