Research map / 3D vision / Face and shape
NILUT: Conditional Neural Implicit 3D Lookup Tables for Image Enhancement
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 NILUT, a coordinate MLP that implicitly represents 3D color LUTs (and a conditional variant covering multiple styles in one network) to make LUT-based image enhancement memory-efficient and mobile-friendly.
Problem
Conventional 3D LUTs require storing a separate ~107K-parameter table per picture style and use operations (e.g., CUDA trilinear interpolation) unsupported on mobile NPUs, limiting their use on mobile devices.
Approach
Represent a 3D color transform as a continuous implicit neural representation (MLP mapping RGB to RGB) fitted to Hald-style RGB maps of real LUTs, and condition a single network on a one-hot or blended style vector so multiple LUTs share one compact model.
Key results
- MLP-Res 128x2 reaches 45.34 dB PSNR and 0.97 delta E on RGB maps, and 42.04 dB PSNR on MIT-5K images.
- A single CNILUT learns five 3D LUTs with an average RGB-map PSNR of 45.34 dB and delta E 0.85 (MLP-Res 256x2).
- An 8.7K-parameter NILUT compresses a 33-dim 3D LUT (~107K parameters) by about 13x; a CNILUT (32x2) runs at ~10 FPS on regular GPUs and processes 4K images on mobile GPUs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MLP-Res instead of a basic MLP | NILUT · RGB map | PSNR_rgb: 43.18 -> 45.34 | Helped |
| Use MLP-Res instead of a basic MLP | NILUT · MIT5K | PSNR_5k: 39.70 -> 42.04 | Helped |
| Use MLP-Res instead of SIREN | NILUT · MIT5K | PSNR_5k: 41.17 -> 42.04 | Helped |
| Increase MLP-Res width from 128 to 256 neurons at 2 layers (ablation) | NILUT · RGB map | PSNR: 45.34 -> 45.84 | Helped |
| Increase MLP-Res depth from 2 to 3 layers at 256 neurons (ablation) | NILUT · RGB map | PSNR: 45.84 -> 46.19 | Helped |
| Increase SIREN depth from 2 to 3 layers at 256 neurons (ablation) | NILUT · RGB map | PSNR: 41.95 -> 43.00 | Helped |
| Increase SIREN depth from 2 to 3 layers at 64 neurons (ablation) | NILUT · RGB map | PSNR: 45.37 -> 43.03 | Hurt |
| Increase MLP depth from 2 to 3 layers at 128 neurons (ablation) | NILUT · RGB map | PSNR: 43.18 -> 42.26 | Hurt |
| Learn three styles with CNILUT rather than one specialized NILUT (ablation) | MLP-Res (256, 2) · RGB map | PSNR: 45.34 -> 45.37 | Mixed |
| Increase CNILUT training capacity from 128x2 to 256x3 (ablation) | CNILUT learning 3 styles · RGB map | PSNR: 43.72 -> 47.26 | Helped |
| Increase CNILUT training capacity from 128x2 to 256x3 (ablation) | CNILUT learning 5 styles · RGB map | PSNR: 42.67 -> 46.05 | Helped |
| Blend three styles using condition weights [0.33, 0.33, 0.33] | CNILUT · MIT5K test images | PSNR: 40.05dB | No clear effect |
Surprising findings
- Blending styles implicitly inside the network matches explicit linear interpolation of processed images (40.05 dB PSNR).
Limitations
- A neural network cannot surpass the efficiency of a direct lookup operation.
- Conditional training requires longer and more complex training with slight performance degradation versus separate NILUTs.
- Evaluated mainly by fitting professional LUTs rather than learning enhancement from data.
Experimental setup
Datasets: MIT-Adobe FiveK professional 3D LUTs (Hald RGB maps)
Benchmarks: MIT-Adobe FiveK (100 RAW images) RGB map fidelity
Metrics: PSNR CIELAB delta E
Baselines: SIREN MLP MLP-Res
Builds on: SIREN Fourier Feature Networks 3D LUT MIT-Adobe FiveK
Closest papers by meaning
- Efficient Neural Network Encoding for 3D Color Lookup Tables (AAAI 2025 · similarity 0.90)
- GLUT: 3D Gaussian Lookup Table for Continuous Color Transformation (NeurIPS 2026 · similarity 0.90)
- Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling (CVPR 2024 · similarity 0.85)
- TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge (NeurIPS 2024 · similarity 0.82)
- Neural 3D Strokes: Creating Stylized 3D Scenes with Vectorized 3D Strokes (CVPR 2024 · similarity 0.82)
- DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · 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.