Research map / Image generation / Image restoration
DeFlow: Learning Complex Image Degradations From Unpaired Data With Conditional Flows
Valentin Wolf, Andreas Lugmayr, Martin Danelljan, Luc Van Gool, Radu Timofte
Check 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 |
|---|---|---|---|
| Use DeFlow-generated degradations instead of prior degradation pipelines | ESRGAN super-resolution training · AIM-RWSR | LPIPS: 0.394 -> 0.349 | Improved |
| Use DeFlow-generated degradations instead of prior degradation pipelines | ESRGAN super-resolution training · NTIRE-RWSR | LPIPS: 0.226 -> 0.218 | Improved |
| Use DeFlow-generated degradations instead of prior degradation pipelines | ESRGAN super-resolution training · DPED-RWSR | BRISQUE: 21.49 -> 21.13 | Improved |
| Increase flow depth from K=4 to K=16 (ablation) | DeFlow · AIM-RWSR | LPIPS: 0.362 -> 0.349 | Improved |
| Increase flow depth from K=8 to K=16 (ablation) | DeFlow · AIM-RWSR | LPIPS: 0.355 -> 0.349 | Improved |
| Use 4× rather than 8× downsampling for h(x) (ablation) | DeFlow · AIM-RWSR | LPIPS: 0.349 -> 0.429 | Hurt |
| Use 16× rather than 8× downsampling for h(x) (ablation) | DeFlow · AIM-RWSR | LPIPS: 0.349 -> 0.352 | Hurt |
| Add domain-invariant conditional information h(x) rather than using h(x)=0 (ablation) | DeFlow · AIM-RWSR | LPIPS: 0.412 -> 0.349 | Improved |
| Learn a correlated rather than non-learned latent shift (ablation) | DeFlow · AIM-RWSR | LPIPS: 0.405 -> 0.349 | Improved |
| Allow correlated rather than uncorrelated learned latent shift (ablation) | DeFlow · AIM-RWSR | LPIPS: 0.359 -> 0.349 | Improved |
| Set the same standard-normal base distribution for both domains (u=0) (ablation) | DeFlow · AIM-RWSR | downstream super-resolution performance | Hurt |
| Use DeFlow instead of CycleGAN for degradation learning | Real-world super-resolution · AIM-RWSR, NTIRE-RWSR, and DPED-RWSR | perceived image quality | Improved |
Later papers that compared with this method
3 later accepted papers report DeFlow as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution (CVPR 2024 · better on 3 of 3 shared results)
- Learning Many-to-Many Mapping for Unpaired Real-World Image Super-Resolution and Downscaling (TPAMI 2024 · better on 1 of 3 shared results)
- Learn From Unpaired Data for Image Restoration: A Variational Bayes Approach (TPAMI 2022 · better on 3 of 6 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
- Reversing Flow for Image Restoration (CVPR 2025 · similarity 0.90)
- Unsupervised Degradation Representation Learning for Unpaired Restoration of Images and Point Clouds (TPAMI 2024 · similarity 0.89)
- Learning Degradation Representations for Image Deblurring (ECCV 2022 · similarity 0.88)
- UniRes: Universal Image Restoration for Complex Degradations (ICCV 2025 · similarity 0.88)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.88)
- Learning from History: Task-agnostic Model Contrastive Learning for Image Restoration (AAAI 2024 · 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.
