Organizations: Tongji Univeristy · Tencent · Nanyang Technological University · Hong Kong University of Science and Technology · University of Hong Kong · National University of Singapore · Fuzhou University
In this paper, we introduce MegaStyle, a novel and scalable data curation pipeline that constructs an intra-style consistent, inter-style diverse and high-quality style dataset. We achieve this by leveraging the consistent text-to-image style mapping capability of current large generative models, which can generate images in the same style from a given style description. Building on this foundation, we curate a diverse and balanced prompt gallery with 170K style prompts and 400K content prompts, and generate a large-scale style dataset MegaStyle-1.4M via content-style prompt combinations. With MegaStyle-1.4M, we propose style-supervised contrastive learning to fine-tune a style encoder MegaStyle-Encoder for extracting expressive, style-specific representations, and we also train a FLUX-based style transfer model MegaStyle-FLUX. Extensive experiments demonstrate the importance of maintaining intra-style consistency, inter-style diversity and high-quality for style dataset, as well as the effectiveness of the proposed MegaStyle-1.4M. Moreover, when trained on MegaStyle-1.4M, MegaStyle-Encoder and MegaStyle-FLUX provide reliable style similarity measurement and generalizable style transfer, making a significant contribution to the style transfer community. More results are available at our project website https://jeoyal.github.io/MegaStyle/.
Figures & tables
Figure 2: Illustrations of (a) artworks by Vincent van Gogh ; (b) style images in OmniStyle-150K generated by SOTA style transfer methods [ 5 , 10 , 55 , 2 , 17 , 60 ] from a reference style image; and (c) images generated by Qwen-Image using the same style description.
Figure 3: Overview of our data curation pipeline. We first collect style and content images from open-source datasets. Next, we apply carefully designed instructions to generate style and content prompts with Qwen3-VL, together with balance sampling. Finally, we use Qwen-Image to generate style images using content-style prompt combinations. Please note that we use simplified content and style prompts for illustrative purposes only .
Figure 4: Visualizations of style reproductions. We first use Qwen3-VL to caption a style prompt from the reference style image, and then generate style reproductions on content–style combinations using Qwen-Image.
Figure 5: Distribution analysis of overall artistic styles in the style prompts. We present the proportions of the top 30 overall artistic styles.
Datasets
Intra-style Consistency
Overall Style
Fine-grained Style
Style Image Number
WikiArt
✗
27
—
80K
JourneyDB
✗
—
300K
4.4M
Style30K
✗
—
1K
30K
IMAGStyle
✓
14
15K
210K
OmniStyle-150K
✓
—
1K
150K
MegaStyle-1.4M
✓
8,355
170K
1.4M
Table 1: Comparison of style datasets. ✓/✗ indicate whether intra-style consistency is provided and — indicates that the statistic is unavailable.
Figure 6: Visualizations of style pairs in MegaStyle-1.4M, where each row shows the same style with different contents.
Figure 7: The architecture of MegaStyle-FLUX.
Figure 8: Visual comparison of top-1 matched style retrieval results between MegaStyle-Encoder, SigLIP, and CSD. The red and green borders indicate incorrect and correct matches, respectively.
mAP@k ↑
Recall@k ↑
Methods
Backbone
1
10
1
10
CLIP
ViT-L
9.29
6.46
9.29
31.56
CSD
ViT-L
45.60
37.78
45.60
79.18
MegaStyle-Encoder
ViT-L
87.26
85.98
87.26
97.61
SigLIP
SoViT
10.43
7.83
10.43
36.32
MegaStyle-Encoder
SoViT
88.46
86.77
88.46
97.66
Table 2: Comparison of MegaStyle-Encoder with other style encoders on the StyleRetrieval benchmark. The best results are highlighted in bold .
Figure 9: Qualitative comparison between MegaStyle-FLUX and SOTA style transfer methods. MegaStyle-FLUX achieves the superior performance compared to baseline methods.
Metrics
StyleCrafter
DEADiff
Attn-Distill
InstantStyle
CSGO
StyleAligned
StyleShot
MegaStyle-FLUX
Style ↑
48.59
51.34
85.59
71.41
55.02
59.80
63.42
76.16
Text ↑
21.39
23.13
20.29
20.77
23.05
21.31
21.79
23.20
Human Style ↑
3.41
3.05
13.97
18.19
7.34
7.46
15.21
31.37
Human Text ↑
8.87
11.13
6.31
10.98
16.18
4.12
13.69
28.72
Table 3: Quantitative comparison of style and text alignment with SOTA style transfer methods. Style and Text denote the cosine distance in the feature spaces of MegaStyle-Encoder and CLIP, respectively. The prefix Human indicates human preference scores. Best result is marked in bold , and the second-best result is highlighted in underline .
Figure 10: Visual results of MegaStyle-FLUX trained on different style datasets.
Figure 11: Visual results of MegaStyle-FLUX and fine-tuned StyleShot.
Metrics
JourneyDB
OmniStyle-150K
MegaStyle-1.4M
Style ↑
34.56
51.49
76.16
Text ↑
21.12
23.02
23.20
Table 4: Quantitative comparison of style datasets. Best is marked in bold .
StyleBench
FLUX-Retrieval
OmniStyle-150K
mAP@k ↑
Recall@k ↑
mAP@k ↑
Recall@k ↑
mAP@k ↑
Recall@k ↑
Methods
1
10
1
10
1
10
1
10
1
10
1
10
CLIP
40.00
30.85
40.00
82.50
2.42
1.55
2.42
9.68
1.68
1.35
1.68
10.39
CSD
70.00
51.65
70.00
97.50
14.16
9.91
14.16
40.08
60.86
48.24
60.86
89.71
MegaStyle-Encoder
85.00
54.15
85.00
100.00
22.70
18.38
22.70
51.87
78.89
60.18
78.89
94.07
Table 5: Comparison of MegaStyle-Encoder with other style encoders on the StyleBench, FLUX-Retrieval and OmniStyle-150K. The best results are highlighted in bold .
Metrics
StyleShot-FLUX
StyleShot-FLUX-Mega
MegaStyle-FLUX
Style ↑
57.06
67.73
76.16
Text ↑
21.86
23.27
23.20
Table 6: Quantitative comparison between MegaStyle-FLUX and fine-tuned StyleShot. Best is marked in bold .
Figure 12: Visualizations of retrieval benchmark WikiArt and our StyleRetrieval, where each row shows the same style during retrieval.
Figure 13: Visual results between MegaStyle-FLUX and Qwen-Image-Edit.
Metrics
Qwen-Image-Edit
MegaStyle-FLUX
Style ↑
43.03
76.16
Text ↑
24.24
23.20
Table 7: Quantitative comparison between MegaStyle-FLUX and Qwen-Image-Edit. Best is marked in bold .
Figure 14: Visualizations of association bias in Qwen-Image.
Figure 15: Additional visualizations of style pairs in MegaStyle-1.4M, where each row shows the same style with different contents.
Figure 16: Additional visualizations of style pairs in MegaStyle-1.4M, where each row shows the same style with different contents.
Figure 17: Additional visualizations of style pairs in MegaStyle-1.4M, where each row shows the same style with different contents.
Figure 18: Additionaly qualitative comparison between MegaStyle-FLUX and SOTA style transfer methods.
Figure 19: Additionaly qualitative comparison between MegaStyle-FLUX and SOTA style transfer methods.
Image style is a highly abstract, human-constructed concept shaped by a range of visual factors and intrinsically entangled with content, yet a unified and explicit definition of image style remains lacking. In this work, we first discuss the fundamental question of what is style and then propose a hierarchical style definition that describes image style from an overall style identity to fine-grained visual attributes, providing a more structured, transferable, and interpretable style representation. Based on this definition, we refine the style annotation pipeline of MegaStyle and construct MegaStyle++-8M, a large-scale style dataset containing 150K overall style identities, 1M fine-grained style prompts, and 8M stylized images. Extensive analyses demonstrate that our hierarchical definition substantially expands the style space in both diversity and semantic breadth, while precisely capturing intrinsic visual style of reference images. The dataset and code will be updated at https://github.com/Tencent/MegaStyle, we hope MegaStyle++ provides a scalable foundation for studying and modeling diverse image styles.
Junyao Gao, Sibo Liu, Jiaxing Li +4
Tongji University · Tencent · Nanyang Technological University +1
Content-Preserving Style transfer, given content and style references, remains challenging for Diffusion Transformers (DiTs) due to entangled content and style features. With a reverse triplet synthesis pipeline to build a million-scale training set and a dual-branch Style-Content DiT (SC-DiT) that decouples style and content via separate ROPE embeddings and causal masking, we observe that such a one-stage training paradigm on mixed style categories causes semantic styles to dominate, hindering texture style learning, and harming content preservation. To address these issues, we propose Style-CCL, a Multi-Stage Curriculum Continual Learning framework that trains SC-DiT from semantic (easy) to texture (hard) styles, and from clean to synthetic data, with Random Memory Rehearsal across stages to avoid catastrophic forgetting. Extensive experiments demonstrate that our Style-CCL achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.
Shiwen Zhang, Haoyuan Wang, Xianghao Zang +3
Institute of Artificial Intelligence (TeleAI), China Telecom.
Tone style transfer for photo retouching aims to adapt the stylistic tone of the reference image to a given content image. However, the lack of high-quality large-scale triplet datasets with stylized ground truth forces existing methods to rely on self-supervised or proxy objectives, which limits model capability. To mitigate this gap, we design a data construction pipeline to build TST100K, a large-scale dataset of 100,000 content-reference-stylized triplets. At the core of this pipeline, we train a tone style scorer to ensure strict stylistic consistency for each triplet. In addition, existing methods typically extract content and reference features independently and then fuse them in a decoder, which may cause semantic loss and lead to inappropriate color transfer and degraded visual aesthetics. Instead, we propose ICTone, a diffusion-based framework that performs tone transfer in an in-context manner by jointly conditioning on both images, leveraging the semantic priors of generative models for semantic-aware transfer. Reward feedback learning using the tone style scorer is further incorporated to improve stylistic fidelity and visual quality. Experiments demonstrate the effectiveness of TST100K, and ICTone achieves state-of-the-art performance on both quantitative metrics and human evaluations.
Yuhai Deng, Huimin She, Wei Shen +4
VCIP, School of Computer Science, Nankai University · OPPO AI Center, OPPO Inc. · Nankai International Advanced Research Institute (Shenzhen Futian)