Research map / Image generation / Image restoration
Look-Up Table Compression for Efficient Image Restoration
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 Diagonal-First Compression (DFC), which splits a spatial LUT into a diagonal re-indexed LUT and a sparsely subsampled non-diagonal LUT to cut LUT storage for image restoration, and introduces SPF-LUT as a stronger multi-LUT model.
Problem
LUT-based image restoration models gain performance from multiple LUTs, but storage grows rapidly and exceeds the on-device cache size, hindering edge deployment.
Approach
Observes a diagonal-dominance property in LQ patch occurrence statistics: most cached entries lie near the diagonal. Diagonal entries are preserved and re-indexed into a lower-dimensional LUT, while non-diagonal entries are subsampled at a large interval, converting one LUT into two much smaller LUTs.
Key results
- DFC reduces SPF-LUT storage from 17.284MB to 2.018MB (compression ratio 11.7%) on x4 SR, losing only 0.06 dB on Set5 (31.11 to 31.05).
- MuLUT +DFC stores 0.407MB vs 4.062MB for original MuLUT, with PSNR on Set5 dropping from 30.60 to 30.55.
- The paper claims DFC saves up to 90% of storage at negligible performance drop for x4 super-resolution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DFC to SPF-LUT | SPF-LUT · Set5 ×4 super-resolution | PSNR: 31.11 -> 31.05 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · Set14 ×4 super-resolution | PSNR: 27.92 -> 27.88 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · BSDS100 ×4 super-resolution | PSNR: 27.10 -> 27.08 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · Urban100 ×4 super-resolution | PSNR: 24.87 -> 24.81 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · Manga109 ×4 super-resolution | PSNR: 28.68 -> 28.58 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · Set12 grayscale denoising, noise level 15 | PSNR: 32.11 -> 32.01 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · BSD68 grayscale denoising, noise level 15 | PSNR: 31.17 -> 31.09 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · Classic5 image deblocking, JPEG quality factor 10 | PSNR-B: 28.63 -> 28.62 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · LIVE1 image deblocking, JPEG quality factor 10 | PSNR-B: 28.62 -> 28.61 | Hurt |
| Apply DFC to SPF-LUT | SPF-LUT · GoPro deblurring | PSNR: 25.94 -> 25.92 | Hurt |
| Increase diagonal width λ from 1 to 5 (ablation) | SPF-LUT +DFC · Set5 ×4 super-resolution | PSNR | Helped |
| Increase diagonal width λ beyond 5 toward 11 (ablation) | SPF-LUT +DFC · Set5 ×4 super-resolution | PSNR | Helped |
Surprising findings
- Performance saturates at diagonal width 9 with about 60% compression ratio, suggesting at least 40% of SPF-LUT storage is redundant.
Limitations
- Gains are evaluated mainly on PSNR/SSIM with LUT-aware finetuning needed to recover performance.
- Compression hyperparameters (p, lambda, sigma) are chosen per model and task; default config tuned for small compression ratio.
- Efficiency and deployment details are deferred to supplementary material.
Experimental setup
Datasets: DIV2K GoPro
Benchmarks: Set5 Set14 BSDS100 Urban100 Manga109 Set12 BSD68 Classic5 LIVE1 GoPro
Metrics: PSNR SSIM PSNR-B storage size compression ratio
Baselines: SR-LUT MuLUT NE + LLE ANR A+ RRDB EDSR BM3D WNNM TNRD DnCNN FFDNet
Builds on: SR-LUT MuLUT IMDN
Closest papers by meaning
- TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge (NeurIPS 2024 · similarity 0.87)
- CoDe: An Explicit Content Decoupling Framework for Image Restoration (CVPR 2024 · similarity 0.83)
- Multi-Frame Deformable Look-Up Table for Compressed Video Quality Enhancement (AAAI 2025 · similarity 0.82)
- Test-Time Degradation Adaptation for Open-Set Image Restoration (ICML 2024 · similarity 0.81)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.81)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.81)
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.