Research map / Image generation / Super-resolution
Toward Real-World Super-Resolution via Adaptive Downsampling Models
Sanghyun Son, Jaeha Kim, Wei‐Sheng Lai, Ming–Hsuan Yang, Kyoung Mu Lee
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 |
|---|---|---|---|
| use LFL data term instead of L1 to bicubic-downsampled HR (ablation) | unsupervised downsampler · x2 k4 LR PSNR | PSNR between LR and Down: 37.22 -> 43.75 | Improved |
| use ADL instead of LFL data term (ablation) | unsupervised downsampler · x4 k0 LR PSNR | PSNR between LR and Down: 31.13 -> 38.24 | Improved |
| train SR on ADL-downsampled images instead of bicubic (ablation) | EDSR · synthetic DIV2K x2 k4 | PSNR: 27.95 -> 32.05 | Improved |
| use ADL instead of LFL for the downsampler (ablation) | EDSR · synthetic DIV2K x2 k4 | PSNR: 31.57 -> 32.05 | Improved |
| use a larger RRDB SR backbone with ADL (ablation) | ADL · synthetic DIV2K x4 k4 | PSNR: 26.97 -> 27.19 | Improved |
| use ADL instead of LFL (ablation) | EDSR · RealSR-V3 Nikon x4 | PSNR: 25.90 -> 26.46 | Improved |
| raise balancing weight alpha from 1 to 100 (ablation) | ADL + EDSR · x2 k4 | PSNR: 29.08 -> 32.05 | Improved |
| raise balancing weight alpha from 1 to 200 (ablation) | LFL + EDSR · x2 k4 | PSNR: 29.36 -> 31.57 | Improved |
| increase training LR images from 1 to 100 (ablation) | LFL + EDSR · x2 k4 | PSNR: 25.43 -> 31.21 | Improved |
| add ADL over LFL with a single LR image (ablation) | EDSR · x2 k4, 1 LR image | PSNR: 25.43 -> 29.84 | Improved |
| backpropagate SR loss into downsampler (end-to-end) (ablation) | ADL + EDSR · x2 k4 | PSNR: 32.09 -> 31.39 | Hurt |
| add fake quantization to end-to-end training (ablation) | ADL + EDSR · x2 k4 | PSNR: 31.39 -> 31.50 | No clear effect |
Closest papers by meaning
- Learning Controllable Degradation for Real-World Super-Resolution via Constrained Flows (ICML 2023 · similarity 0.92)
- 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)
- Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts (CVPR 2024 · similarity 0.88)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.88)
- Navigating Beyond Dropout: An Intriguing Solution towards Generalizable Image Super Resolution (CVPR 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.
