We present LoR-LUT, a unified low-rank formulation for compact and interpretable 3D lookup table (LUT) generation. Unlike conventional 3D-LUT-based techniques that rely on fusion of basis LUTs, which are usually dense tensors, our unified approach extends the current framework by jointly using residual corrections, which are in fact low-rank tensors, together with a set of basis LUTs. The approach described here improves the existing perceptual quality of an image, which is primarily due to the technique's novel use of residual corrections. At the same time, we achieve the same level of trilinear interpolation complexity, using a significantly smaller number of network, residual corrections, and LUT parameters. The experimental results obtained from LoR-LUT, which is trained on the MIT-Adobe FiveK dataset, reproduce expert-level retouching characteristics with high perceptual fidelity and a sub-megabyte model size. Furthermore, we introduce an interactive visualization tool, termed LoR-LUT Viewer, which transforms an input image into the LUT-adjusted output image, via a number of slidebars that control different parameters. The tool provides an effective way to enhance interpretability and user confidence in the visual results. Overall, our proposed formulation offers a compact, interpretable, and efficient direction for future LUT-based image enhancement and style transfer.
Figures & tables
Figure 1: Comparison between conventional adaptive 3D-LUTs and our LoR-LUT. (a) conventional adaptive 3D-LUTs rely on fusing several dense basis LUTs. (b) LoR-LUT replaces dense fusion with a compact low-rank residual, achieving similar flexibility with fewer parameters and identical interpolation complexity.
Figure 2: Trilinear interpolation in a 3D LUT. Each input color (r,g,b) lies within a cube defined by eight lattice vertices L[xi,yj,zk] . The output color y is a weighted sum of eight vertices of the cube, where wijk is a trilinear interpolation weight that decreases with distance from (r,g,b) along each axis.
Figure 3: Low-rank residual representation. Each rank-1 component factorizes the correction tensor into three axis-specific vectors and a color coefficient. Summing R such components yields a compact yet expressive residual tensor ΔL .
Figure 4: Overall architecture of LoR-LUT. Two lightweight networks predict the fusion weights and low-rank residual factors from the input image. The reconstructed LUT is then applied using the standard trilinear interpolation operator. The entire pipeline is differentiable and end-to-end trainable.
Module
Configuration
Output
Weight predictor
Conv( 3→16 ), Conv( 16→32 ), AvgPool, FC( 32→K )
α∈RK
Residual predictor
Conv( 3→16 ), Conv( 16→32 ), AvgPool, FC ×4
R×(3G+3)
Basis LUTs (if K>0 )
K learnable tensors of size G3×3
K×3G3
Table 1: Network architecture details. Layers used in the weight and residual predictors of LoR-LUT, together with the learnable basis LUTs when K>0 . Here G is the LUT grid size and R is the low-rank decomposition rank.
Figure 5: Visualization of low-rank residual correction in LoR-LUT. (a) Base LUT (identity, K=0 ); (b) residual output; and (c) combined base+residual result. The bottom row shows the corresponding 3D LUT cubes. The low-rank residual introduces subtle yet structured color shifts.
Method
Params
Interp.
IA-3D-LUT [ 15 ]
0.59M
Trilinear
AdaInt [ 12 ]
0.61M
Adaptive
SA-3D-LUT [ 11 ]
4.5M
Trilinear
LoR-LUT (ours)
37K
Trilinear
Table 2: Model size comparison. Number of learnable parameters for representative image-adaptive 3D-LUT methods. All methods use a LUT grid size of G=33 .
Method
PSNR ↑
SSIM ↑
LPIPS ↓
HDRNet [ 4 ]
22.15
0.840
0.182
DPE [ 2 ]
23.75
0.908
0.094
IA-3D-LUT [ 15 ]
22.27
0.837
0.183
AdaInt [ 12 ]
25.13
0.921
-
SepLUT [ 13 ]
25.32
0.918
-
LoR-LUT
Table 3: Quantitative comparison on MIT-Adobe FiveK (Expert C). LoR-LUT achieves the best perceptual quality
Figure 6: Qualitative comparison on MIT-Adobe FiveK (Expert C). From left to right: input image, HDRNet [ 4 ] , CSRNet [ 5 ] , CLUT [ 16 ] , IA-3DLUT [ 15 ] , our residual-only variant (LoR-LUT, K=0 ), our full model (LoR-LUT), and the expert ground truth (GT). Compared to prior LUT-based and CNN-based baselines, LoR-LUT produces more natural color tones, smoother local contrast, and better overall perceptual fidelity, closely matching the expert edit.
R
PSNR ↑
SSIM ↑
LPIPS ↓
ΔE00↓
Params
4
25.13
0.895
0.081
6.16
0.024M
8
25.35
0.901
0.079
5.85
0.037M
32
25.53
0.901
0.083
5.75
0.118M
Table 4: Effect of the low-rank R for the residual-only LoR-LUT ( K=0 ).
Method
Params
Model size
4K latency (T4)
LoR-LUT
(K=0,R=32)
0.12M
0.45MB
68.3ms
LoR-LUT
(K=8, R=32)
0.98M
3.74MB
69.0ms
LoR-LUT
(K=8, R=0)
0.87M
3.33MB
68.1ms
Table 5: Runtime and parameter analysis of LoR-LUT variants on 4K inputs (3840 × 2160). Latency is averaged over 100 runs on NVIDIA T4 and L4 GPUs. All models use grid size G=33 .
Figure 7: LoR-LUT Viewer for exploring low-rank residuals. The web-based tool takes an input image, displays the enhanced output and the corresponding 3D LUT cube, and exposes sliders for the magnitude of each rank-1 component.
Figure 8: Low-rank residual factors on an example image. Top: for each rank-1 component, we plot its three 1D factor curves along the R, G, and B axes of the LUT grid, revealing separable tonal trends such as highlight boosts or shadow roll-off. Bottom: the corresponding RGB coefficients indicate how strongly each component acts on each color channel. Together, these factors make individual components interpretable as ”color brushes” in the 3D LUT cube.