Research map / Image generation / Super-resolution
AutoLUT: LUT-Based Image Super-Resolution with Automatic Sampling and Adaptive Residual Learning
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 AutoLUT, a LUT-based image super-resolution framework with learnable automatic sampling (AutoSample) and adaptive residual learning (AdaRL) as plug-and-play modules that enlarge receptive fields and improve PSNR while cutting storage and inference time.
Problem
LUT-based super-resolution networks rely on fixed, hand-designed sampling patterns that limit receptive field flexibility and fine-detail capture, and cannot use residual connections without LUT value overflow, so accuracy and efficiency are constrained for edge devices.
Approach
Replace fixed LUT sampling patterns with a learnable softmax-normalized convolution that selects weighted pixel combinations (AutoSample, expanding receptive field via pixel shuffle), and add a learnable convex-combination residual (AdaRL) whose output is guaranteed to stay in the [0,255] LUT range.
Key results
- MuLUT + AutoLUT improves average PSNR by about 0.20 dB over MuLUT across five benchmarks (Set5 30.60 to 30.85 dB) with storage nearly unchanged (4.062 MB to 4.067 MB).
- SPF-Light (one branch per group) cuts storage by 55% versus SPF-LUT + DFC while keeping comparable PSNR (31.02 vs 31.05 dB on Set5).
- On a mobile device, MuLUT + Ours runs in 5984 ms versus 5938 ms for MuLUT (similar time) while gaining 0.25 dB on Set5; AutoLUT reduces inference time by about two-thirds for MuLUT and SPF-LUT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AutoSample and AdaRL to MuLUT | MuLUT · Set5, Set14, BSD100, Urban100, and Manga109 ×4 | PSNR: 30.60 -> 30.85; 27.60 -> 27.77; 26.86 -> 26.96; 24.46 -> 24.60; 27.90 -> 28.27 | Helped |
| Reduce SPF-LUT branch count to create SPF-Light | SPF-LUT · Set5 ×4 mobile | runtime: 31921.65 -> 9910.95 | Helped |
| Reduce SPF-LUT branch count to create SPF-Light | SPF-LUT · Set5 ×4 | PSNR: 31.05 -> 31.02 | Hurt |
| Use one branch and sample size 5 | MuLUT · Set5 ×4 mobile | runtime: 5938.1 -> 2043.95 | Helped |
| Use one branch and sample size 5 | MuLUT · Set5 ×4 mobile | PSNR: 30.60 -> 30.62 | Helped |
| Increase sample size from 3 to 5 (ablation) | MuLUT with 3 branches · Set5 ×4 | PSNR: 30.79 -> 30.85 | Helped |
| Increase sample size from 5 to 7 (ablation) | MuLUT with 3 branches · Set5 ×4 | PSNR: 30.85 -> 30.82 | Hurt |
| Increase branch count from 2 to 3, with sample size 3 (ablation) | MuLUT · Set5 ×4 | PSNR: 30.73 -> 30.79 | Helped |
| Increase branch count from 3 to 4, with sample size 3 (ablation) | MuLUT · Set5 ×4 | PSNR: 30.79 -> 30.83 | Helped |
| Add AutoSample (ablation) | MuLUT · Set5 ×4 | PSNR: 30.60 -> 30.63 | Helped |
| Add AdaRL (ablation) | MuLUT · Set5 ×4 | PSNR: 30.60 -> 30.70 | Helped |
| Add AutoSample and AdaRL (ablation) | MuLUT · Set5 ×4 | PSNR: 30.60 -> 30.79 | Helped |
Surprising findings
- Increasing sample size from 5 to 7 slightly decreased PSNR (by 0.03 dB on Set5), attributed to irrelevant information from an overly large receptive field.
- AdaRL alone gives a larger gain (30.70 dB) than AutoSample alone (30.63 dB) on Set5 in the ablation.
Limitations
- Gains on MuLUT are small (about 0.2 dB), and the sample-size ablation shows diminishing or negative returns beyond size 5.
- Storage increases with the number of branches, so higher-capacity configurations trade off against the efficiency goal.
Experimental setup
Datasets: DIV2K
Benchmarks: Set5 Set14 BSDS100 Urban100 Manga109
Metrics: PSNR SSIM storage size inference time
Baselines: SR-LUT MuLUT SPF-LUT SPF-LUT + DFC NE+LLE ANR A+ RRDB EDSR RCAN SwinIR Bicubic
Builds on: SR-LUT MuLUT SPF-LUT rotation ensemble DFC LUT compression
Closest papers by meaning
- Adaptive Dropout: Unleashing Dropout across Layers for Generalizable Image Super-Resolution (CVPR 2025 · similarity 0.85)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.83)
- Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance (CVPR 2024 · similarity 0.83)
- Auto-Encoded Supervision for Perceptual Image Super-Resolution (CVPR 2025 · similarity 0.82)
- Text-guided Explorable Image Super-resolution (CVPR 2024 · similarity 0.82)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.82)
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.