Research map / Image generation / Super-resolution
To learn image super-resolution, use a GAN to learn how to do image degradation first
Adrian Bulat, Jing Yang, Georgios Tzimiropoulos
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 |
|---|---|---|---|
| train Low-to-High on High-to-Low (learned degradation) output instead of bilinear downsampling (ablation) | Low-to-High · Widerface real-world LR test set | FID: 85.59 -> 14.89 | Improved |
| train Low-to-High on High-to-Low output instead of bilinear plus blur (ablation) | Low-to-High · Widerface real-world LR test set | FID: 84.68 -> 14.89 | Improved |
| add GAN loss to pixel loss when training Low-to-High on High-to-Low output (ablation) | Low-to-High · Widerface real-world LR test set | FID: 87.91 -> 14.89 | Improved |
| use learned High-to-Low degradation instead of bilinear downsampling to create LR images | degradation model · Widerface real-world LR test set | FID to original LR: 23.15 -> 15.27 | Improved |
| add blurring before bilinear downsampling for training LR images (ablation) | Low-to-High · Widerface real-world LR test set | FID: 85.59 -> 84.68 | No clear effect |
| compare against SRGAN trained on bilinear pairs | SRGAN · Widerface real-world LR test set | FID: 104.80 -> 14.89 | Improved |
| compare against CycleGan | CycleGan · Widerface real-world LR test set | FID: 19.01 -> 14.89 | Improved |
| train Low-to-High on High-to-Low output instead of bilinear downsampling (ablation) | Low-to-High · LS3D-W (bilinear input) | PSNR: 23.50 -> 19.30 | Hurt |
| train with pixel loss plus GAN loss instead of pixel loss only on LS3D-W bilinear input (ablation) | Low-to-High · LS3D-W (bilinear input) | PSNR: 23.22 -> 19.30 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- FSRNet — FSRNet: End-to-End Learning Face Super-Resolution With Facial Priors (CVPR 2018) · better on 1 of 2 shared results
- Wavelet-SRNet — Wavelet-SRNet: A Wavelet-Based CNN for Multi-Scale Face Super Resolution (ICCV 2017) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning to Super-Resolve Blurry Face and Text Images (ICCV 2017 · similarity 0.93)
- Unpaired Image Super-Resolution Using Pseudo-Supervision (CVPR 2020 · similarity 0.91)
- Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (CVPR 2017 · similarity 0.91)
- SRFeat: Single Image Super-Resolution with Feature Discrimination (ECCV 2018 · similarity 0.91)
- Details or Artifacts: A Locally Discriminative Learning Approach to Realistic Image Super-Resolution (CVPR 2022 · similarity 0.90)
- Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts (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.
