Photographic color grading shapes the tone and color of a photograph to achieve a desired look. Text-driven photographic look transfer lets users specify these adjustments in natural language, with the goal of reproducing the requested look across photographs while preserving their content. However, prior research on text-driven color grading uses different datasets, text instructions, target definitions, and metrics. These differences make it difficult to compare how faithfully the methods reproduce the same look. To compare methods under a shared protocol, we introduce LUTBench. We collect 91 three-dimensional lookup tables (3D LUTs) and define one photographic look per LUT with a specified input conversion. Applying these looks to 5,000 MIT-Adobe FiveK photographs produces 455,000 input-target image pairs with reproducible target images. For each look, we use templates to turn measured tone and color changes on fixed color samples into candidate clauses. We use training image measurements to verify applicable clauses before composing the text description. The evaluation measures target fidelity, tests sensitivity to predefined input conversions, and separates generalization to unseen looks from reproduction of seen looks on new photographs. Across 16 methods, stronger text alignment does not guarantee target fidelity. However, most supervised variants improve fidelity on both seen and unseen looks.
Figures & tables
Figure 2: Overview of LUTBench. LUTs render paired target images. Measurements guide clause generation and verification to construct look descriptions. The paired images and descriptions support evaluation of target fidelity and semantic alignment.
Figure 3: Color and tone variation across 91 looks. Panel A marks the median pairwise mean CIEDE2000. Dots denote looks, boxes show medians and interquartile ranges, and whiskers span the full range. Dashed lines mark unchanged attribute values.
Figure 4: Each point represents one of 91 looks, measured on the same 497 fixed color samples. Point colors encode mean ΔL∗ . Backgrounds indicate chromatic directions in A and smoothed local lightness changes in B.
Figure 5: The main bar chart shows look-group counts by size and partition. The inset boxplot compares pairwise target color differences within groups, between groups, and across partitions.
Canonical ΔE00↓
CV fidelity
Regional ΔE00↓
Weighted components
Text alignment ↑
Method
CV (40)
Val (9)
Test (10)
Seen (72)
Min. error ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Neutral
Chrom.
Hue-bal.
L (bias)
C (bias)
H
Clauses
CLIP dir.
Canonical target image
0.00
0.00
0.00
0.00
0.00
∞
1.000
0.000
0.00
0.00
0.00
0.0 (0.0)
0.0 (0.0)
0.0
94%
0.036
Unedited input
8.65
9.38
10.26
7.66
7.78
22.62
0.873
0.109
8.11
9.84
10.04
5.7 (-4.5)
3.1 (-1.7)
3.7
7%
-
CLIP-guided tone
CLIPtone [ 17 ]
13.18 +
13.80 +
12.77
14.08 +
12.57
21.64
0.773
0.185
12.87
13.34
13.39
4.8 (-0.6)
6.9 (+5.5)
8.2
41%
0.099
LCIT [ 23 ]
10.10 +
10.77 +
11.15 +
9.08 +
9.30
22.18
0.861
0.144
8.62
13.47
12.80
5.8 (-4.6)
3.4 (-1.1)
5.6
19%
0.083
Table 1: Target fidelity and text alignment. Blue headers denote the Seen track, and red headers denote the Unseen track (CV, Val and Test). Regional errors, weighted components and text-alignment metrics use CV looks. Header counts indicate the number of looks. ∗ marks configurations supervised on LUTBench. In the canonical-error columns, - and + indicate supported improvements and deficits relative to the unedited input under the rule in Sec. 4.3 , respectively. Unmarked method comparisons are inconclusive. Weighted components report mean absolute weighted CIEDE2000 terms, with signed lightness and chroma biases in parentheses. Min. error, Clauses and CLIP dir. denote minimum error over candidate input conversions, clause satisfaction and CLIP directional similarity, respectively.
Figure 6: Same held-out look and description across three photographs. Each three-row block aligns Input, the canonical target and all 22 method configurations in a fixed order. The shared description is below. Labels give mean pixelwise ΔE00↓ to the canonical target. Input denotes the unedited input. ∗ denotes LUTBench supervision.
Figure 7: Panel A plots raw CLIP directional similarity against canonical error. Panel B subtracts category means from both metrics. Both show 21 configurations. Dashed lines show linear fits, through the origin in B. r is Pearson correlation.
Figure 8: Clause satisfaction versus target fidelity. Lightness changes ΔL∗ and output/input chroma ratios are medians over neutral midtone and blue pixels identified in the full input image, respectively. Quoted clauses require ΔL∗≥1.5 and ratio ≤0.8 . Boxes mark enlarged regions.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Conversion category
Looks
Median largest shift
Calibrated log
45
2.59
Display
28
3.87
Uncalibrated log
15
4.59
Cross-encoding alternatives
3
7.43
Appendix
Table S1 : Sensitivity to the LUT input conversion, measured by the difference between canonical and alternative target images on 500 validation photographs. The last column gives the median, across looks, of the largest mean ΔE00 over alternatives.
Group
Clause
Holds when
Uses
Tone
“neutral { zone } rendered { slightly, clearly, much } { brighter, darker } ”
magnitude of median ΔL∗ for neutral pixels in [2,5] , [5,12] , or [12,∞) , with sign matching brighter/darker. Directional variant: ≥1.5
129
“a lowered white point”
white sample lightness L∗≤95 (fixed samples)
49
“a lifted black point”
black sample lightness L∗≥4 (fixed samples)
25
“ { soft, pronounced } contrast along the neutral tone curve”
contrast ratio ≤0.85 or ≥1.15 (fixed samples)
36
Color cast
“neutral { zone } rendered with a [ { subtle, clear, strong } ] { color } cast”
magnitude of the zone-median a∗ or b∗ in [2,5] , [5,12] , or [12,∞) , with sign matching the color cast. Directional variant: ≥1.5
Table S2 : Candidate-clause templates and numerical conditions. Uses counts clause occurrences across the 91 descriptions. Conditions marked fixed samples use fixed color samples, and the other conditions use training photographs. The lightness change ΔL∗ is measured from input to canonical target image, and color casts use a∗ and b∗ in the canonical target image at initially neutral pixels. Within each hue sector, the chroma ratio is the median ratio of CIELAB chroma in the canonical target image to that in the input photograph.
Work
Implementation and Model
Data and Evaluation Material
Evaluated Methods
CLIPtone [ 17 ]
Code and adapter weights
FiveK and Color Names. Nine-filter rendered pairs: NI.
LCIT [ 23 ]
Paper-based implementation. Author code: NI.
Kodak images and prompt examples.
CLIPstyler [ 16 ]
Code and example decoders
DIV2K, CelebA-HQ and repository examples.
FastCLIPstyler [ 31 ]
Code and weights
Painter by Numbers, DTD and ArtEmis. Exact evaluation data: NI.
SpectralCLIP [ 37 ]
Code using pretrained CLIP
COCO images and style names. Exact pair assignments: NI.
Appendix
Table S3 : Implementation and data availability of the evaluated methods and of related methods that LUTBench does not evaluate. Headings of the related-method groups give the reason for exclusion. The data column distinguishes source datasets from exact evaluation material. NI denotes material not identified in the inspected release as of November 2026. A dash indicates no separate dataset entry. Work names link to the cited release or paper.
Method minus unedited input
CV
Group CI
Evid. (32)
Lower sensitivity (21)
Both criteria (13)
Test
CLIPtone [ 17 ]
+4.53 +
[+3.20,+5.91]
+4.51 +
+4.26 +
+4.06 +
+2.50
LCIT [ 23 ]
+1.45 +
[+1.24,+1.66]
+1.45 +
+1.55 +
+1.61 +
+0.89 +
CLIPstyler [ 16 ]
+13.99 +
[+12.36,+15.75]
+13.78 +
+13.72 +
+13.03 +
+13.42 +
FastCLIPstyler [ 31 ]
+22.99 +
[+20.20,+26.08]
+22.41 +
+23.67 +
+22.65 +
+21.06 +
SpectralCLIP [ 37 ]
+21.55 +
[+19.89,+23.21]
+21.66 +
+21.62 +
+21.93 +
+20.30 +
ZeCon [ 39 ]
+6.28 +
[+5.82,+6.67]
+6.30 +
+6.33 +
+6.41 +
+5.46 +
Appendix
Table S4 : Sensitivity of paired differences relative to the unedited input on the corresponding evaluation photograph sets. Positive values indicate higher canonical error. Group CI is the 95% confidence interval for the difference in canonical error obtained by resampling complete look groups. The CV input-conversion subsets use explicit evidence for LUT input encoding (Evid.), largest mean differences between canonical and alternative target images below 4 ΔE00 (Lower sensitivity), or both conditions (Both criteria). Superscripts + / − mark supported differences. Paired differences use unrounded values.
Method
Canonical error ΔE00↓
PSNR ↑
SSIM ↑
LPIPS ↓
Difference from unedited input
Supervised minus zero-shot
Unedited input
7.66 [7.13,8.20]
23.54
0.865
0.110
-
-
CLIPtone [ 17 ]
14.08 [13.32,14.81]
20.81
0.777
0.195
+6.42 + [+5.60,+7.24]
-
LCIT [ 23 ]
9.08 [8.62,9.56]
23.13
0.856
0.144
+1.42 + [+1.31,+1.53]
-
CLIPstyler [ 16 ]
22.22 [21.36,23.16]
15.43
0.428
0.447
+14.56 + [+13.57,+15.67]
-
FastCLIPstyler [ 31 ]
32.18 [30.65,33.70]
11.41
0.483
0.549
+24.53 + [+22.66,+26.28]
-
SpectralCLIP [ 37 ]
30.98 [30.33,31.66]
10.68
0.225
0.580
+23.32 + [+22.36,+24.25]
-
Appendix
Table S5 : Fidelity and paired comparisons in the Seen track on 72 training looks and all 500 test photographs. Brackets give 95% confidence intervals. Differences are method minus the unedited input or supervised minus the corresponding zero-shot configuration. Superscripts + / − mark differences supported by both canonical and minimum error. ∗ denotes supervised variants. The evaluation scope is defined in Sec. A4 .
Rel.
Adapter Target ∗
Adapter CLIP
All Scratch ∗
All Fine-tune ∗
Canon. ΔE00↓
13.18
6.81
16.08
6.38
7.69
CI lower
12.08
6.13
15.14
5.74
6.87
CI upper
14.32
7.54
16.99
7.06
8.60
Min. error ↓
12.57
6.39
15.62
6.04
6.97
PSNR ↑ (dB)
21.64
26.17
20.10
26.85
25.19
SSIM ↑
0.773
0.907
0.728
0.925
0.902
Appendix
Table S6 : CLIPtone [ 17 ] training configurations on 40 CV looks and 500 validation photographs. Rel. is the released model. Adapter columns update only the text adapter, using target or CLIP supervision. Columns labeled All update both the enhancer and text adapter, from scratch or by fine-tuning. CI rows give the 95% confidence interval bounds for canonical error. ∗ marks target supervision. Training settings are in Sec. B2 .
Figure S1 : Human evaluation interface. Input-target pairs and the full look description appear on the left, with five-point clause ratings and a content-skip option on the right.
Measure
Result
95% CI
Denom.
Mean score (1–5)
3.93
[3.86,4.00]
91/8,208
Scores ≥4 (%)
68.94
[65.78,72.10]
91/8,208
Scores =5 (%)
32.79
[29.92,35.93]
91/8,208
No relevant content (%)
15.38
[13.32,17.36]
91/9,765
All three scores ≥4 (%)
45.11
[40.33,50.05]
91/2,262
Ordinal α
0.595
[0.568,0.621]
1,672/8,164
Appendix
Table S7 : Human evaluation of look descriptions. Denom. gives eligible looks/responses, looks/clause-rater units for the three-photograph measure, and items/ratings for α . Score measures exclude content skips. Skip rates include all responses.
Figure S2 : Canonical look examples on eight landscape training photographs. Each row shows the input and six LUT-rendered target images, using matched 4:3 crops. Full look descriptions appear below the grid, keyed by color and look identifier.
Figure S3 : Canonical look examples on five portrait-oriented training photographs and eight further looks. Each row uses matched 3:4 crops. Full look descriptions appear below the grid. Both photographs and looks differ from those in Fig. S2 .
Figure S4 : Single-photograph comparison under Look S032. The full description appears at left. The grid shows Input, the canonical target and all 22 method configurations.