ColorFLUX: A Structure-Color Decoupling Framework for Old Photo Colorization

Bingchen Li; Zhixin Wang; Fan Li; Jiaqi Xu; Jiaming Guo; Renjing Pei; Xin Li; Zhibo Chen

cs.CVMar 30, 2026

ColorFLUX: A Structure-Color Decoupling Framework for Old Photo Colorization

Authors: Bingchen Li, Zhixin Wang, Fan Li, Jiaqi Xu, Jiaming Guo, Renjing Pei, Xin Li, Zhibo Chen

Organizations: University of Science and Technology of China · Huawei Noah’s Ark Lab

Abstract

Old photos preserve invaluable historical memories, making their restoration and colorization highly desirable. While existing restoration models can address some degradation issues like denoising and scratch removal, they often struggle with accurate colorization. This limitation arises from the unique degradation inherent in old photos, such as faded brightness and altered color hues, which are different from modern photo distributions, creating a substantial domain gap during colorization. In this paper, we propose a novel old photo colorization framework based on the generative diffusion model FLUX. Our approach introduces a structure-color decoupling strategy that separates structure preservation from color restoration, enabling accurate colorization of old photos while maintaining structural consistency. We further enhance the model with a progressive Direct Preference Optimization (Pro-DPO) strategy, which allows the model to learn subtle color preferences through coarse-to-fine transitions in color augmentation. Additionally, we address the limitations of text-based prompts by introducing visual semantic prompts, which extract fine-grained semantic information directly from old photos, helping to eliminate the color bias inherent in old photos. Experimental results on both synthetic and real datasets demonstrate that our approach outperforms existing state-of-the-art colorization methods, including closed-source commercial models, producing high-quality and vivid colorization.

Figures & tables

Figure 2: llustration of real old photo’s fading effect. ColorFLUX brings them back to life (second row).

Figure 3: The framework of our proposed method. (a) The inference process of colorization, including a preprocessor to remove degradations in old photos. (b) The training procedure of our method. Notably, we illustrate an example of easy-to-hard losing samples. The model progressively learns preferences from severely augmented samples to milder ones, ensuring precise color perception to effectively counteract the fading effects of old photos. (c) The preference pair collection pipeline. To synthesize $x^{l}$ , we randomly apply augmentation combinations sampled from [‘B’, ‘C’, ‘S’].

DIV2K-valid-synthesized
	NR-IQA			Qwen-score						Runtime
	DeQA	Q-Insight	VQ-R1	CRI	CRA	CCS	SCS	AES	OA	Runtime
DeOldify [ 2 ]	4.080	3.978	4.625	69.35	80.95	83.80	94.78	80.70	81.73	2.23s
DDColor [ 19 ]	4.162	3.992	4.574	79.80	78.55	77.60	93.22	81.13	81.80	0.19s
CtrlColor [ 27 ]	4.144	4.013	4.654	75.40	80.25	84.04	94.53	81.85	83.28	8.05s
FLUX-Kontext [ 24 ]	4.209	4.002	4.661	77.75	80.00	83.61	93.91	82.22	83.48	28.14s

Table 1: Quantitative comparisons across different methods on two benchmarks. VQ-R1, CRI, CRA, CCS, SCS, AES, and OA denote VisualQuality-R1, Color RIchness, Color RAtionality, Color ConSistency, Structure ConSistency, AESthetics, and OverAll performance, respectively. Notably, we apply our preprocessor to all methods to ensure a fair comparisons on colorization performance. The Runtime is measured on a single GPU with about 19.5 TFLOPS. Best results are bolded.

Figure 4: Qualitative comparisons between ColorFLUX and other methods on RealOldPhotos . Zoom in for better views.

Figure 5: Visual results of decoupling each parts of ColorFLUX. We denote structure consistency training, basic color learning, and fine-color adjustment by SCT, BCL, and FCA, respectively. “W/o” represents that we remove the according network from the trained framework ( e.g . , w/o SCT denotes for removing ControlNet).

	DeQA	Q-Insight	VQ-R1	CRI	CRA	CCS	AES
DIV2K-valid-augmented
Text Prompt	3.758	3.689	4.289	50.00	68.70	78.82	69.74
VS Prompt	3.932	3.752	4.344	55.80	71.85	79.44	72.24
RealOldPhotos
Text Prompt	3.863	3.761	4.248	58.00	71.70	75.00	70.06
VS Prompt	4.105	3.811	4.381	67.80	79.60	82.16	78.62

Table 2: Quantitative comparisons across different semantic prompt strategies. “VS Prompt” refers to visual semantic prompt. Best results are bolded.

Figure 6: Qualitative comparisons for different fine-color post training strategy on DIV2K-valid-augmented .

	DeQA	Q-Insight	VQ-R1	CRI	CRA	AES	OA
$\mathcal{L}_{\text{FM}}$	3.889	3.868	4.432	46.35	68.45	67.64	69.91
$\mathcal{L}_{\text{distill}}$	4.044	3.895	4.552	58.55	70.65	72.04	75.12
$\mathcal{L}_{\text{FM}}+\mathcal{L}_{\text{distill}}$	4.114	3.907	4.577	57.55	72.45	73.32	75.39

Table 3: Quantitative comparisons across different basic color learning strategies. “

\mathcal{L}_{\text{FM}}

” and “

\mathcal{L}_{\text{distill}}

” denote flow matching loss and distillation loss, respectively. We report results on DIV2K-valid-synthesized without augmentations.

Figure 7: Visual comparisons between different basic color learning strategies on DIV2K-valid-augmented .

	DeQA	Q-Insight	VQ-R1	CRI	CRA	CCS	AES
DIV2K-valid-augmented
W/o DPO	3.932	3.752	4.344	55.80	71.85	79.44	72.24
SFT	4.236	3.939	4.656	70.95	78.65	84.02	79.80
One-stage DPO	4.307	3.972	4.692	77.05	79.35	84.15	82.46
Pro-DPO	4.303	3.978	4.728	80.15	79.70	82.66	82.80
RealOldPhotos

Table 4: Quantitative comparisons across different fine-color adjusting strategies. “W/o DPO” indicates that no color adjustment training, and “One-stage DPO” refers to DPO training without the progressive strategy. Best results are bolded.

Figure 8: User study of ColorFLUX vs. other methods on RealOldPhotos among 10 participants.

Figure 9: A screenshot of our user study.

Figure 10: Visual comparisons between ColorFLUX and closed-source GPT-4o. Zoom in for better views. GPT results exhibit noticeable structural inconsistencies, particularly in regions such as fingers and faces.

Figure 11: The template of our proposed Qwen-score.

Figure 12: Visual comparisons between ColorFLUX and other methods on DIV2K-valid-synthesized dataset.

Figure 13: Visual comparisons between ColorFLUX and other methods on DIV2K-valid-synthesized dataset.

Figure 14: Visual comparisons between ColorFLUX and other methods on DIV2K-valid-augmented dataset.

Figure 15: Visual comparisons between ColorFLUX and other methods on RealOldPhotos dataset.

Figure 16: Colorization results of ColorFLUX on RealOldPhotos .

Figure 17: Colorization results of ColorFLUX on RealOldPhotos .

Figure 18: Colorization results of ColorFLUX on RealOldPhotos .

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 372 papers

Reviving Ancient Paintings via Poem: A Colorization Framework for Aligning Cultural Semantics
Jul 20, 2026Junming Gao, Biao Zhu, Xiaosong Wang +1Diffusion Models for Image Restoration Image-to-Image Translation
Zhejiang University, Hangzhou, China · Hangzhou Research Institute of AI and Holographic Technology, Hangzhou, China
Loop-Mamba: A Loop Mamba with Degradation-Aware and Shared Memory for Old Photo Restoration
Aug 3, 2026Runci Bai, Yucheng Xin, Pu Wang +8Visual SSMs Image Restoration
China Academy of Information and Communications Technology · Shandong Normal University · Shandong University +5
NumColor: Precise Numeric Color Control in Text-to-Image Generation
Mar 13, 2026Muhammad Atif Butt, Diego Hernandez, Alexandra Gomez-Villa +3Controllable Image Generation T2I Diffusion Models
Computer Vision Center, Spain · Computer Sciences Department, Universitat Autònoma de Barcelona, Spain · Program of Computer Science, City University of Hong Kong (Dongguan) +1

Jul 20, 2026J/K move · Enter open · S save

Reviving Ancient Paintings via Poem: A Colorization Framework for Aligning Cultural Semantics

Junming Gao, Biao Zhu, Xiaosong Wang, Tan Tang

Zhejiang University, Hangzhou, China · Hangzhou Research Institute of AI and Holographic Technology, Hangzhou, China

Diffusion Models for Image Restoration Image-to-Image Translation Image Restoration

Open paper AskPDF arXiv

The irreversible fading of ancient paintings disrupts the "congruence between poems and paintings", a core aesthetic principle where visual imagery harmonizes with literary inscriptions. Although diffusion models provide strong generative priors, restoring historically faithful colors remains difficult: visual restoration is inherently ambiguous, while direct text guidance often causes modern semantic bias, over-saturation, and cross-boundary color leakage. To address this, we propose PoemColor, a poem-guided ancient painting colorization framework. Our method aligns poetic cultural semantics with painting restoration through two key designs. First, the Poetic Painting Projector (P3) converts implicit poetic context into a classical color-aware condition via poem-to-palette pretraining, reducing the ambiguity of poem-to-color mapping. Second, Structure-Aware Semantic Attention (SASA) regulates how poetic color semantics are injected into the diffusion backbone by jointly controlling their propagation direction and regional injection strength. In addition, we construct a hybrid restoration dataset that integrates synthetic degradation with expert-restored artifacts, providing both scalable supervision and real classical color references. Extensive experiments demonstrate that our framework significantly outperforms state-of-the-art methods, delivering controllable colorization that revives both historical authenticity and poetic semantics.

Abstract

Image Restoration

Explore similar work

Reviving Ancient Paintings via Poem: A Colorization Framework for Aligning Cultural Semantics

Loop-Mamba: A Loop Mamba with Degradation-Aware and Shared Memory for Old Photo Restoration

NumColor: Precise Numeric Color Control in Text-to-Image Generation