Research map / Image generation / Image restoration
On-Demand Learning for Deep Image Restoration
Ruohan Gao, Kristen Grauman
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 |
|---|---|---|---|
| replace rigid joint learning with on-demand allocation of sub-tasks (ablation) | image restoration network · CelebA Image Deblurring | mean PSNR: 29.40 dB -> 29.58 dB | Improved |
| replace rigid joint learning with on-demand allocation of sub-tasks (ablation) | image restoration network · CelebA Image Inpainting | mean PSNR: 32.11 dB -> 32.30 dB | Improved |
| replace rigid joint learning with on-demand allocation of sub-tasks (ablation) | image restoration network · SUN397 Image Deblurring | mean PSNR: 28.53 dB -> 28.70 dB | Improved |
| replace rigid joint learning with on-demand allocation of sub-tasks (ablation) | image restoration network · SUN397 Image Denoising | mean PSNR: 25.69 dB -> 25.80 dB | Improved |
| replace hard mining with on-demand learning (ablation) | image restoration network · CelebA Image Deblurring | mean PSNR: 27.33 dB -> 29.58 dB | Improved |
| use staged curriculum instead of cumulative joint training (ablation) | image restoration network · CelebA Image Deblurring | mean PSNR: 29.31 dB -> 15.59 dB | Hurt |
| reverse curriculum order (staged anti-curriculum) versus joint learning (ablation) | image restoration network · CelebA Image Deblurring | mean PSNR: 29.40 dB -> 25.43 dB | Hurt |
| use cumulative curriculum instead of rigid joint learning (ablation) | image restoration network · CelebA Image Deblurring | mean PSNR: 29.40 dB -> 28.70 dB | Hurt |
| add 100,000 extra training images to rigid joint learning (200,000 total) (ablation) | rigid joint learning · CelebA | PSNR/L2 | No clear effect |
| replace center-only Pathak inpainter with on-demand inpainter on arbitrary blocks | inpainting · CelebA arbitrary square block | PSNR: 11.80 dB -> 20.31 dB | Improved |
| on-demand all-rounder versus Pathak center inpainter on central blocks | inpainting · CelebA central square block | PSNR: 22.16 dB -> 20.74 dB | Hurt |
| train on a single corruption level (fixated) instead of all levels (ablation) | fixated model · CelebA, all difficulty levels | qualitative restoration quality | Hurt |
Closest papers by meaning
- Revitalizing Convolutional Network for Image Restoration (TPAMI 2024 · similarity 0.91)
- Deep Image Prior (CVPR 2018 · similarity 0.89)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.87)
- Stochastic Deep Restoration Priors for Imaging Inverse Problems (ICML 2025 · similarity 0.87)
- Boosting Image Restoration via Priors from Pre-trained Models (CVPR 2024 · similarity 0.87)
- Omni-Kernel Network 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.
