Research map / 3D vision / Depth and stereo
Efficient Neural Network Encoding for 3D Color Lookup Tables
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 a compact invertible-capable residual MLP (ENNeLUT) that embeds 512 3D color LUTs in under 0.25 MB with mean ΔE ≤ 2.0, optionally trained on natural-image color distributions for ΔE ≤ 1.0.
Problem
Storing hundreds of 3D color lookup tables (LUTs) consumes large memory (over 100 MB for 512 zipped LUTs), and existing methods compress individual LUTs rather than exploiting similarities across a library.
Approach
A neural network of residual MLP blocks conditioned on a LUT index via learned per-block embedding biases, with tanh-inverse/tanh wrapping, activation normalization, small initialization near identity, and spectral-normalized residual blocks giving an optional bijective (invertible) variant.
Key results
- Medium model (235 KB) embedding 512 LUTs reaches mean ΔE 2.00 on Hald images and 2.37 on natural images; large model (313 KB) reaches 1.64 and 1.98.
- Training on natural-image color distribution with a ΔE loss gives mean ΔE 0.96 on natural images for the medium model, at the cost of much worse Hald-image error (6.04).
- Under the paper's ablation, the proposed 12.9K-parameter model gets mean ΔE 1.07 on natural images vs 1.83 for MLP-Res from Conde et al. with the same parameter count.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with natural-image sampling instead of uniform sampling (ablation) | Medium model embedding 512 LUTs · Natural images | mean ΔE: $2.44{\scriptstyle \pm0.03}$ -> $1.09{\scriptstyle \pm0.01}$ | Helped |
| Train with natural-image sampling instead of uniform sampling (ablation) | Medium model embedding 512 LUTs · 256³ Hald image | mean ΔE: $2.01{\scriptstyle \pm0.02}$ -> $6.63{\scriptstyle \pm0.12}$ | Hurt |
| Train with ΔE loss instead of L2 loss, using uniform sampling (ablation) | Medium model embedding 512 LUTs · 256³ Hald image | mean ΔE: $2.01{\scriptstyle \pm0.02}$ -> $1.65{\scriptstyle \pm0.02}$ | Helped |
| Train with ΔE loss instead of L2 loss, using uniform sampling (ablation) | Medium model embedding 512 LUTs · Natural images | mean ΔE: $2.44{\scriptstyle \pm0.03}$ -> $1.92{\scriptstyle \pm0.01}$ | Helped |
| Apply proposed initialization and normalization to MLP-Res (ablation) | MLP-Res embedding 32 LUTs · 256³ Hald image | mean ΔE: $1.54{\scriptstyle \pm0.19}$ -> $1.16{\scriptstyle \pm0.32}$ | Helped |
| Replace MLP-Res with the proposed medium architecture (ablation) | Model with proposed inductive biases embedding 32 LUTs · 256³ Hald image | mean ΔE: $1.16{\scriptstyle \pm0.32}$ -> $0.83{\scriptstyle \pm0.09}$ | Helped |
| Replace MLP-Res with the proposed medium architecture (ablation) | Model with proposed inductive biases embedding 32 LUTs · Natural images | mean ΔE: $1.41{\scriptstyle \pm0.34}$ -> $1.07{\scriptstyle \pm0.12}$ | Helped |
| Evaluate on all MIT5K images instead of 100 random images (ablation) | Medium model trained uniformly on 32 LUTs · Natural images | mean ΔE: $1.3809$ -> $1.4025$ | No clear effect |
Surprising findings
- Training on natural-image colors with a ΔE loss improves natural-image error but raises Hald-image mean ΔE to about 6, showing a strong trade-off with full-gamut accuracy.
- Restricting the network to be strictly bijective requires roughly twice the memory footprint to match the unrestricted model's LUT fidelity.
Limitations
- Strictly bijective variant needs more depth and parameters to match capacity of the unrestricted model.
- Applying the model per pixel to full images is slow (about 239 ms for 12MP on RTX4090), so it is intended for LUT extraction rather than image processing.
- Natural-image-trained models can degrade on rare colors and out-of-distribution colors.
Experimental setup
Datasets: 543 Creative Commons 3D LUTs Adobe-MIT5K
Benchmarks: 256^3 Hald images Adobe-MIT5K natural images
Metrics: mean delta E (CIE76) 90th-percentile delta E PSNR compression ratio
Baselines: MLP-Res (Conde et al.) zip compression of .cube files direct inverse LUT estimation
Builds on: residual networks residual flows activation normalization spectral normalization Conde et al. neural LUT embedding
Closest papers by meaning
- NILUT: Conditional Neural Implicit 3D Lookup Tables for Image Enhancement (AAAI 2024 · similarity 0.90)
- GLUT: 3D Gaussian Lookup Table for Continuous Color Transformation (NeurIPS 2026 · similarity 0.90)
- GamutMLP: A Lightweight MLP for Color Loss Recovery (CVPR 2023 · similarity 0.86)
- How Hard Is It for Message-Passing GNNs to Simulate One Weisfeiler-Lehman Color-Refinement Step? (NeurIPS 2026 · similarity 0.82)
- Learning Color Equivariant Representations (ICLR 2025 · similarity 0.82)
- Color-Oriented Redundancy Reduction in Dataset Distillation (NeurIPS 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.