Research map / 3D vision / Gaussian splatting
GLUT: 3D Gaussian Lookup Table for Continuous Color Transformation
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 GLUT, an explicit continuous 3D LUT representation built from learnable 3D Gaussian primitives with local affine color transforms, plus CGLUT, a conditional generator that encodes many LUT styles for blending and a bias-only local editing scheme.
Problem
Grid-based 3D LUTs suffer from a capacity-memory trade-off, while implicit neural LUT representations (INRs) are black-box, hard to interpret, and cannot be locally edited without global retraining.
Approach
Model a color transformation as a soft partition of RGB space by normalized opacity-weighted Gaussian densities, each applying a local affine color transform plus a global affine residual; a conditional MLP generator maps style embeddings to Gaussian parameters, and edits adjust only the biases of the top-K most influential Gaussians.
Key results
- GLUT-64 reaches 48.42 PSNR on 75-LUT Hald images vs 38.84 for NILUT (128x2) with 1,420 params vs 33,923.
- CGLUT-64 (Large) achieves 52.41 PSNR on 225-LUT Hald images, outperforming CNILUT (256x3) at 41.46 and ENNELUT (L2) at 40.68.
- GLUT-16 reaches 241 FPS on GPU at 4K with CUDA vs 18 FPS for NILUT (64x2).
Surprising findings
- Removing the residual connection (pure local mapping) drops GLUT-32 PSNR from 45.47 to 40.54, showing residual learning matters a lot.
- CGLUT on CPU is slower than ENNELUT and NILUT in FPS, and CNILUT runs below 0.01 FPS at 4K on CPU in the reported measurements.
Limitations
- Editing is not designed for color-specific or semantically aware manipulation, only mood/style adjustment.
- Performance on natural images trained from a single image pair is lower because natural images cover sparse parts of RGB space.
- CPU runtime is slower than traditional LUTs and some neural baselines; efficiency depends on GPU/NPU acceleration.
Experimental setup
Datasets: 300 collected open-source LUTs (.cube files) Hald images MIT-Adobe FiveK
Benchmarks: Hald image dataset (75/225 LUTs) MIT-Adobe FiveK natural images (100 test images)
Metrics: PSNR SSIM LPIPS CIEDE2000 CIE76 Delta E GFLOPs FPS parameter count compression ratio
Baselines: NILUT CNILUT ENNELUT INRetouch
Builds on: 3D Gaussian Splatting NILUT CNILUT ENNELUT
Closest papers by meaning
- NILUT: Conditional Neural Implicit 3D Lookup Tables for Image Enhancement (AAAI 2024 · similarity 0.90)
- Efficient Neural Network Encoding for 3D Color Lookup Tables (AAAI 2025 · similarity 0.90)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (CVPR 2025 · similarity 0.85)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.85)
- Learning Color Equivariant Representations (ICLR 2025 · similarity 0.85)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 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.