Research map / Image generation / Super-resolution
Rethinking Image Super Resolution from Training Data Perspectives
Go Ohtani, Ryu Tadokoro, Ryosuke Yamada, Yuki M Asano, Iro Laina, Christian Rupprecht, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka, Yoshimitsu Aoki
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 SR models on DiverSeg-I (low-resolution filtered ImageNet) instead of DF2K | HAT · Urban100 x4 | PSNR (dB): 27.93 -> 28.51 | Improved |
| train on DiverSeg-I instead of DF2K | RCAN · Urban100 x4 | PSNR (dB): 26.73 -> 27.03 | Improved |
| train on DiverSeg-I instead of DF2K | MSRResNet · Urban100 x4 | PSNR (dB): 26.23 -> 26.31 | Improved |
| train on DiverSeg instead of ImageNet pretraining followed by DF2K fine-tuning | HAT · Urban100 x4 | PSNR (dB): 28.37 -> 28.51 | Improved |
| filter by SAM segment count instead of blockiness (ablation) | HAT on ImageNet-1k subset · Urban100 x4 | PSNR (dB): 28.39 -> 28.51 | Improved |
| use segmentation-based rather than detection-based filtering (ablation) | HAT on ImageNet-1k subset · Urban100 x4 | PSNR (dB): 28.44 -> 28.51 | Improved |
| apply segment threshold theta=100 versus no filtering (ablation) | HAT on ImageNet-1k · Urban100 x4 | PSNR (dB): 28.41 -> 28.51 | Improved |
| raise segment threshold from 100 to 150 (ablation) | HAT on ImageNet-1k · Urban100 x4 | PSNR (dB): 28.51 -> 28.36 | Hurt |
| train on JPEG-compressed DF2K (quality 50) instead of HR (ablation) | HAT · Urban100 x4 | PSNR (dB): 27.93 -> 25.04 | Hurt |
| apply object-based filtering to LSDIR (ablation) | HAT · Urban100 x4 | PSNR (dB): 28.45 -> 28.53 | Improved |
| reduce blockiness threshold from 30 to 10 (ablation) | HAT on ImageNet-1k · Urban100 x4 | PSNR (dB): 28.43 -> 28.41 | No clear effect |
| train on Places365 (low estimated quality) instead of ImageNet-1k or PASS (ablation) | HAT · learning curves | PSNR | Hurt |
Closest papers by meaning
- Bridging the Perception Gap in Image Super-Resolution Evaluation (CVPR 2026 · similarity 0.91)
- Seven Ways to Improve Example-Based Single Image Super Resolution (CVPR 2016 · similarity 0.90)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.89)
- SEAL: A Framework for Systematic Evaluation of Real-World Super-Resolution (ICLR 2024 · similarity 0.89)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.89)
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (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.
