Research map / Image generation / Super-resolution
WARP: Wide Attention with Rich Projections for Image Super-Resolution
Jiwon Kim, 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 |
|---|---|---|---|
| Add nonlinear shared QKV projection instead of linear QKV (same depth) (ablation) | WARP C=96 · Set5 x4 SR | PSNR (dB): 32.18 -> 32.48 | Improved |
| Add nonlinear projection versus linear QKV with depth increased to L=9 (ablation) | WARP C=96 · Set5 x4 SR | PSNR (dB): 32.47 -> 32.48 | No clear effect |
| Share QKV projection across blocks instead of independent projections (ablation) | WARP C=96 · Set5 x4 SR | PSNR (dB): 32.41 -> 32.48 | Improved |
| Add channel attention branch to spatial-only attention (ablation) | WARP C=96 · Set5 x4 SR | PSNR (dB): 32.43 -> 32.48 | Improved |
| Add nonlinear projection instead of linear QKV with L=20 at full scale (ablation) | WARP C=320 · Urban100 x4 SR | PSNR (dB): 27.65 -> 27.81 | Improved |
| Add dual attention (L=12) instead of spatial-only with L=16 at full scale (ablation) | WARP C=320 · Urban100 x4 SR | PSNR (dB): 27.81 -> 27.95 | Hurt |
| Enlarge attention window from 16x16 to 64x64 (trained from scratch) (ablation) | WARP C=320 · Urban100 x4 SR | PSNR (dB): 26.84 -> 28.08 | Improved |
| Enlarge window from 16x16 to 64x64 on smooth Set5 (ablation) | WARP C=320 · Set5 x4 SR | PSNR (dB): 32.65 -> 33.13 | Improved |
| Use naive tile 128 extrapolation instead of training tile 64 (ablation) | WARP x2 · Urban100 x2 SR | PSNR (dB): 34.85 -> 34.72 | Hurt |
| Add standard YaRN temperature (tau~1.03) to naive extrapolation (ablation) | WARP x2 · Urban100 x2 SR | PSNR (dB): 34.72 -> 34.58 | Hurt |
| Use inverted temperature tau=0.80 instead of standard YaRN temperature (ablation) | WARP x2 · Urban100 x2 SR | PSNR (dB): 34.58 -> 34.95 | Improved |
| Add 50K iterations YaRN fine-tuning (ablation) | WARP x2 · Urban100 x2 SR | PSNR (dB): 34.95 -> 34.99 | Improved |
Closest papers by meaning
- LinearSR: Unlocking Linear Attention for Stable and Efficient Image Super-Resolution (ICLR 2026 · similarity 0.90)
- Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping (ICLR 2026 · similarity 0.87)
- SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers (NeurIPS 2026 · similarity 0.87)
- SRSR: Enhancing Semantic Accuracy in Real-World Image Super-Resolution with Spatially Re-Focused Text-Conditioning (NeurIPS 2025 · similarity 0.87)
- Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution (AAAI 2025 · similarity 0.87)
- Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment (NeurIPS 2025 · 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.
