Research map / Image generation / Image restoration
ShiftLUT: Spatial Shift Enhanced Look-Up Tables for Efficient Image Restoration
Xiaolong Zeng, Yitong Yu, Shiyao Xiong, Jinhua Hao, Ming Sun, Chao Zhou, Bin Wang
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 Learnable Spatial Shift (LSS) module (ablation) | ShiftLUT with varying Shift-Block counts and channel numbers · Set5 x4 | PSNR (dB): +0.37 to +0.46 across block counts, +0.30 to +0.43 across channels | Improved |
| add LSS to widen receptive field (ablation) | ShiftLUT · LAM visualization of 16x16 patch | detail index (DI): 3.71 -> 6.13 | Improved |
| make the LSB branch a single 3x3 conv (asymmetric dual branch) instead of a symmetric deep LSB branch (ablation) | ShiftLUT · average of five x4 SR benchmarks | PSNR / latency: PSNR 28.19/0.8016 -> 28.19/0.8014; latency 164ms -> 84ms | Improved |
| use Error-bounded Adaptive Sampling (eps 0.4) to compress LUTs (ablation) | ShiftLUT-L · Set5 x4 | LUT size / PSNR: 217KB -> 104KB at PSNR 31.33 unchanged | Improved |
| increase EAS tolerance from 0.4 to 0.8 (ablation) | ShiftLUT-L · Set5 x4 | LUT size / PSNR: 104KB -> 54KB; PSNR 31.33 -> 31.30 | Mixed |
| use uniform sampling (step 2) instead of EAS (ablation) | ShiftLUT-L · Set5 x4 | runtime (ms): 84 -> 213 | Hurt |
| use DDM compression instead of EAS (ablation) | ShiftLUT-L · Set5 x4 | PSNR / size / runtime: PSNR 31.33 -> 31.26; size 172KB vs 104KB; runtime 98ms vs 84ms | Hurt |
| assign more bits to the MSB branch (MSB/LSB split) (ablation) | ShiftLUT · x4 SR | PSNR and LUT storage | Mixed |
| use a more complex offset prediction network in LSS (ablation) | ShiftLUT · x4 SR | PSNR | No clear effect |
| replace learned continuous offsets with fixed integer shifts at inference (two-stage training) (ablation) | ShiftLUT LSS · inference | inference cost | Improved |
| use ShiftLUT-L instead of TinyLUT-F on x4 SR | LUT-based SISR · Manga109 x4 | PSNR (dB): 28.83 -> 29.16 | Improved |
| use ShiftLUT-L instead of TinyLUT-F for denoising | LUT-based denoising · Set12, BSD68 sigma 15 | PSNR (dB): 32.22 -> 32.43; 31.20 -> 31.32 | Improved |
Closest papers by meaning
- TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge (NeurIPS 2024 · similarity 0.90)
- Toward DNN of LUTs: Learning Efficient Image Restoration With Multiple Look-Up Tables (TPAMI 2024 · similarity 0.90)
- Improving the Learning Capability of Small-size Image Restoration Network by Deep Fourier Shifting (NeurIPS 2024 · similarity 0.90)
- Look-Up Table Compression for Efficient Image Restoration (CVPR 2024 · similarity 0.86)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.86)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.85)
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.
