Research map / Image generation / Super-resolution
CAMixerSR: Only Details Need More "Attention"
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CAMixer, a content-aware token mixer that routes image windows to deformable window self-attention or depthwise convolution via a learned predictor, and builds CAMixerSR for large-image, lightweight, and omnidirectional SR.
Problem
Large-image (2K-8K) super-resolution is computationally expensive; existing content-aware routing accelerators (e.g., ClassSR, ARM) use inflexible patch-level partitioning with limited receptive fields, while lightweight token-mixer designs ignore that different regions need different processing com
Approach
A content-aware mixer that uses a learned predictor to generate offsets for deformable window warping, a hard/simple window mask, and convolutional channel/spatial attentions; high-complexity windows get window self-attention, plain windows get convolution, and a global ratio loss controls the self-attention fraction.
Key results
- On lightweight SR (×4 Urban100), CAMixerSR reaches 26.63 dB at 765K params and 53.8G MAdds, outperforming SwinIR-light (26.47 dB) and SwinIR-NG (26.54 dB at 1201K params).
- On large-image SR, CAMixerSR-B (765K params, 1.96G FLOPs) reaches 27.89 dB on Test4K, matching RCAN (15.6M params, 32.6G FLOPs) at 27.89 dB; the CAMixer variant saves about 25% FLOPs.
- On ODI-SR ×2, CAMixerSR (1.30M params) reaches 30.46 dB PSNR, slightly above OSRT-light (30.42 dB) and well above EDSR (30.32 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CAMixerSR instead of SwinIR-light | Lightweight SR · Urban100 ×4 | PSNR: 26.47 -> 26.63 | Helped |
| use CAMixerSR instead of SwinIR-light | Lightweight SR · Set14 ×4 | PSNR: 28.77 -> 28.82 | Mixed |
| add deformable offsets (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.60 -> 26.63 | Helped |
| add convolutional spatial attention (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.46 -> 26.63 | Helped |
| add convolutional channel attention (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.50 -> 26.63 | Helped |
| remove offsets, spatial attention, and channel attention (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.63 -> 26.45 | Hurt |
| use window size 8×8 instead of 16×16 (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.63 -> 26.41 | Hurt |
| use window size 32×32 instead of 16×16 (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.63 -> 26.74 | Helped |
| add global input condition (ablation) | CAMixer predictor · Urban100 ×4 | PSNR: 26.46 -> 26.54 | Helped |
| add window input condition (ablation) | CAMixer predictor · Urban100 ×4 | PSNR: 26.46 -> 26.55 | Helped |
| set offset scalar r to 8 rather than 0 (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.60 -> 26.63 | Helped |
| set offset scalar r to 16 rather than 8 (ablation) | CAMixerSR · Urban100 ×4 | PSNR: 26.63 -> 26.51 | Hurt |
Surprising findings
- Simple convolution alone performs nearly as well as convolution plus self-attention on simple patches (Table 1: 43.73 vs 43.80 dB on simple patches).
- Cropping images into smaller tiles reduces PSNR for all tested models, showing that patch-based accelerators limit receptive field (Table 2).
Limitations
- ClassSR applied to CAMixerSR falls 0.12 dB behind content-aware routing on F2K, attributed to limited receptive field of small crops (left for future work).
- Larger window sizes (32x32) suffer larger drops at gamma=0.5 because large windows are hard to classify.
Experimental setup
Datasets: DIV2K ODI-SR Flickr2K DIV8K Set5 Set14 BSD100 Urban100 Manga109 SUN360
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109 Test2K Test4K Test8K F2K ODI-SR SUN360
Metrics: PSNR SSIM WS-PSNR WS-SSIM FLOPs
Baselines: ClassSR ARM IMDN SwinIR-light ELAN-light NGswin SwinIR-NG RCAN OSRT-light EDSR
Builds on: SwinIR ClassSR ARM ACMix Gumbel softmax
Closest papers by meaning
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.85)
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (CVPR 2024 · similarity 0.85)
- WaveMixSR-V2: Enhancing Super-resolution with Higher Efficiency (Student Abstract) (AAAI 2025 · similarity 0.85)
- Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution (AAAI 2025 · similarity 0.84)
- CATANet: Efficient Content-Aware Token Aggregation for Lightweight Image Super-Resolution (CVPR 2025 · similarity 0.84)
- SeD: Semantic-Aware Discriminator for Image Super-Resolution (CVPR 2024 · similarity 0.84)
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.