GenColor: Generative and Expressive Color Enhancement with Pixel-Perfect Texture Preservation

Neurips 2025 (Spotlight)

Yi Dong^1*, Yuxi Wang^1*, Xianhui Lin², Wenqi Ouyang¹, Zhiqi Shen^1†, Peiran Ren^2†, Ruoxi Fan¹, Rynson W. H. Lau^3†

¹Nanyang Technological University, ²Alibaba Group, ³City University of Hong Kong

*Equal contribution †Corresponding authors

Explore the comprehensive navigation below to access detailed comparisons and analyses of GenColor against various baseline, expert, and state-of-the-art methods. Each section provides visual and textual insights into the strengths and unique features of GenColor for generative and expressive color enhancement. Due to space limitations in the main paper, the complete FreeRaw quantitative results are provided here.

Supplementary Video

A. Baseline Color Enhacement Comparison (Adobe5K, PPR10K, FreeRaw)

A1. Input vs GenColor

This section presents a direct visual comparison between the original input images and the results produced by GenColor. Observe how GenColor enhances color vibrancy and detail while preserving the natural appearance of each photo.

A2. 3D-LUT vs GenColor

Here, we compare GenColor with the 3D-LUT method, highlighting GenColor's ability to achieve more natural and visually pleasing enhancements, especially in challenging lighting and color conditions.

A3. RSFNet vs GenColor

This section showcases the performance of GenColor versus RSFNet. Notice how GenColor maintains texture fidelity and color accuracy, outperforming RSFNet in challenging scenarios.

A4. DeepLPF vs GenColor

DeepLPF is a strong baseline for local parametric filtering. Here, we demonstrate how GenColor surpasses DeepLPF by delivering more consistent color enhancement.

A5. ICELUT vs GenColor

ICELUT is compared with GenColor to illustrate the advantages of our approach in handling complex scenes and maintaining both color and texture integrity.

A6. Distort & Restore vs GenColor

Distort & Restore (D&R) is a reinforcement learning-based enhancement method. This section highlights how GenColor provides more reliable and visually appealing results, especially in difficult cases.

A7. Exposure vs GenColor

Exposure is a white-box photo post-processing framework. Here, we show how GenColor achieves superior color balance and detail retention compared to Exposure.

B. Human Expert Comparison (Adobe5K, PPR10K)

B1. Input vs GenColor

In this section, we compare the input images with GenColor outputs, using Human Expert C as the ground truth. This highlights GenColor's ability to closely match expert-level enhancements.

B2. Human Expert C vs GenColor

This section provides a direct comparison between Human Expert C and GenColor, emphasizing GenColor's effectiveness in replicating expert adjustments and producing high-quality results.

C. Texture Preservation Module Removes Artifacts

C1. Diffusion Generated Reference vs GenColor

GenColor leverages a texture preservation module to effectively remove artifacts that are present in diffusion-generated reference images. In the examples below, the left image shows the result from a diffusion model, which often contains visible artifacts and texture inconsistencies. The right image demonstrates how GenColor refines these results, producing artifact-free outputs with natural textures and faithful color reproduction. This highlights the importance of texture preservation for high-quality color transfer.

D. Texture Preservation Comparison vs Color Transfer Methods

D1. NLUT vs GenColor

NLUT is a color transfer method that often fails to preserve the reference image's color fidelity, resulting in significant color deviations. GenColor, in contrast, maintains accurate color reproduction and natural appearance.

D2. Color Matcher vs GenColor

Color Matcher operates on global color statistics, which can lead to noticeable color mismatches and loss of local detail. GenColor excels by providing both global and local color consistency, ensuring visually pleasing results.

D3. UniColor vs GenColor

UniColor offers patch-level color guidance but lacks fine-grained control and cannot adjust image lightness. GenColor provides more precise and flexible color enhancement, resulting in superior visual quality.

D4. Deep Preset vs GenColor

Deep Preset is limited in its ability to preserve the reference image's colors, often causing large deviations. GenColor, however, ensures faithful color transfer and robust texture preservation.

E. Texture Preservation Comparison vs Style Transfer Methods

E1. StyA2K vs GenColor

StyA2K, a style transfer method, struggles with texture preservation and may introduce unnatural halo artifacts. GenColor maintains both texture detail and natural color transitions, delivering more realistic results.

E2. CAP-VSTNet vs GenColor

CAP-VSTNet can introduce unnatural color and halo artifacts during style transfer. GenColor avoids these issues, producing artifact-free, visually coherent images with preserved textures.

F. FreeRaw Dataset and Quantitative Results

In this section, we provide the FreeRaw dataset and the quantitative results of GenColor. Please refer to the main PDF for the quantitative results on Adobe-FiveK and PPR10K Datasets.

F1. Visual Examples

If you would like to view the FreeRaw dataset images used in our visual examples, please visit the following website: https://www.signatureedits.com/free-raw-photos/.

F2. Quantitative Results

Quantitative Results Table: Table below presents a comprehensive comparison of GenColor and various state-of-the-art and baseline methods on the FreeRaw dataset, evaluated using five key color enhancement quality metrics. Each metric is introduced with its corresponding number for clarity:

Q-Align (↑): A non-reference perceptual quality metric that evaluates image aesthetic quality without requiring expert-retouched references, considering both color and structural aspects.
LAION (↑): An aesthetic score from the LAION database, reflecting how well the enhanced result aligns with its dataset’s learned aesthetic preferences.
LIQE (↑): A no-reference image quality assessment framework that employs multi-task learning to jointly optimize scene classification and distortion identification, enabling robust perceptual quality prediction without reference images.
NoR-VDP (↑): A No-Reference Visual Difference Predictor designed for high dynamic range content, assessing image quality without requiring a reference image, focusing on natural image statistics, dynamic range perception, and human visual system response to luminance variations.
C-VAR (↑): A measure capturing the “expressiveness” or color variation introduced by the enhancement. Higher values indicate more vibrant or diverse transformations.

Interpretation: For all metrics, higher values indicate better performance. As shown in Table, GenColor (Ours) achieves the best results across all five evaluation metrics, with the following values: Q-Align↑: 4.73, LAION↑: 6.13, LIQE↑: 4.38, NoR-VDP↑: 68.12, C-VAR↑: 13.27. This demonstrates GenColor's superior performance in color enhancement compared to existing methods.

Method	FreeRaw
Method	Q-Align↑	LAION↑	LIQE↑	NoR-VDP↑	C-VAR↑
3D-LUT	4.44	5.93	4.02	64.14	10.26
RSFNet	4.49	5.92	4.14	64.75	8.64
DeepLPF	4.35	5.93	3.99	65.98	11.74
ICELUT	4.33	5.86	3.84	63.48	8.45
D&R	2.53	5.19	1.29	64.34	6.61
D&R(ARTISAN)	2.42	5.09	1.26	63.90	6.18
Exposure	4.13	5.90	3.44	62.57	13.70
Exposure(ARTISAN)	4.05	5.82	3.37	62.21	13.21
GenColor	4.51	5.96	4.22	67.03	14.93

Reference

Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning Image-Adaptive 3D Lookup Tables for High Performance Photo Enhancement in Real-Time. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 4, pages 2058–2073, 2022. [3D-LUT]
Jongchan Park, Joon-Young Lee, Donggeun Yoo, and In So Kweon. Distort-and-Recover: Color Enhancement Using Deep Reinforcement Learning. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). [D&R]
Sidi Yang et al. Taming Lookup Tables for Efficient Image Retouching. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. [ICELUT]
Jiaqi Ouyang, Shijie Li, Shuai Li, Wangmeng Zuo, and Shuhang Gu. RSFNet: Rich Feature Attention Network for Image Enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. [RSFNet]
Yuanming Hu, Hao He, Chenxi Xu, Baoyuan Wang, and Stephen Lin. Exposure: A White-Box Photo Post-Processing Framework. ACM Trans. Graph., vol. 37, no. 2, May 2018. [Exposure]
Sean Moran, Pierre Marza, Steven McDonagh, Sarah Parisot, and Gregory Slabaugh. DeepLPF: Deep Local Parametric Filters for Image Enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. [DeepLPF]
Christopher Hahne and Amar Aggoun. PlenoptiCam v1.0: A Light-Field Imaging Framework. IEEE Transactions on Image Processing, vol. 30, pages 6757-6771, 2021. [Color Matcher]
Zhitong Huang, Nanxuan Zhao, and Jing Liao. UniColor: A Unified Framework for Multi-Modal Colorization with Transformer. ACM Transactions on Graphics, vol. 41, no. 6, pages 1-16, 2022. [UniColor]
Linfeng Wen, Chengying Gao, and Changqing Zou. CAP-VSTNet: Content Affinity Preserved Versatile Style Transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2023, pages 18300-18309. [CAP-VSTNet]
Zhanghan Ke, Yuhao Liu, Lei Zhu, Nanxuan Zhao, and Rynson W.H. Lau. Color Matcher for Color Style Transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2023, pages 14173-14182. [NeuralPreset]
Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Frédo Durand. Learning Photographic Global Tonal Adjustment with a Database of Input/Output Image Pairs. In The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition, 2011. [Adobe-FiveK]
Jie Liang, Hui Zeng, Miaomiao Cui, Xuansong Xie, and Lei Zhang. PPR10K: A Large-Scale Portrait Photo Retouching Dataset with Human-Region Mask and Group-Level Consistency. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. [PPR10K]
Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. [ControlNet]