SuperStyle · High-Fidelity Large-Scale Style Transfer Dataset

OverviewWhat is SuperStyle?

SuperStyle is a large-scale, high-quality paired style transfer dataset that unifies two complementary lines of work into a single resource that is both broad in style coverage and faithful in supervision:

Breadth · OmniStyle

Stylization-driven synthesis

Large-scale style-transfer data synthesized by stylizing content with multiple SOTA models.

Content–Style–Stylized triplets
1,000+ style categories
Unified instruction-based evaluation & filtering
OmniStyle, CVPR 2025

Fidelity · OmniStyle2

Destylization-driven synthesis

High-fidelity data built by destylizing real artworks, using them as the only ground truth.

Destylized–Style–Target triplets
Real artworks as authentic supervision
CoT-reinforced interpretable quality filtering
OmniStyle2 / DeStyle, ECCV 2026

Together they yield 500K+ triplets spanning 1000+ styles, providing a scalable and reliable data foundation for controllable, generalizable, high-fidelity style transfer.

SuperStyle shifts style transfer toward a supervised learning paradigm — combining style breadth with supervision fidelity at scale.

01Background & Motivation

Image style transfer renders one image (paintings, cartoons, line art, …) in the artistic style of another, while preserving the original content and structure. It is widely used in digital art, advertising creativity and film production.

The field has evolved from early CNN-based single-image optimization (Gatys et al., CVPR 2016) to modern pipelines centered on diffusion-model fine-tuning (e.g., SigStyle, AAAI 2025). Yet two key challenges remain:

1) Insufficient style generalization: existing methods handle only a few simple styles (oil painting, sketch, watercolor) and struggle to generalize to more diverse styles.

2) Poor transfer fidelity: current style-transfer data are low in quality, making high-fidelity style transfer hard to achieve.

Therefore, building a large-scale, high-quality paired dataset with both style breadth and supervision fidelity is key to breaking through current technical bottlenecks — and is exactly what SuperStyle delivers.

02·ABreadth Data — Stylization-driven (OmniStyle)

Goal: build a large-scale, high-quality style-transfer triplet dataset to support controllable and generalizable model training.

Content x Style stylization matrix — Figure A1. Stylization-driven construction: each content image is transferred across diverse styles to form Content–Style–Stylized triplets.

Data construction pipeline

1) Content generation: ChatGPT produces 20 content categories (animals, buildings, humans, …); FLUX generates 2,000 high-resolution content images.

2) Style selection: 1,000 style images curated from Style30K, covering cartoon, watercolor, Chinese painting and many more.

3) Stylized generation: 6 mainstream methods (StyleID, StyleShot, CSGO, ArtFlow, AesPANet, CAST) produce 6 stylized images per content–style pair — yielding ~1M triplets.

OmniStyle-1M instruction-based stylization examples — Figure A2. Instruction-based stylization examples from OmniStyle-1M.

Unified quality evaluation & filtering

A multimodal large model scores each stylized result from three angles:

Content preservation: CLIP for semantic consistency, DINOv2 for structural consistency.
Style consistency: a contrastive task fine-tunes CLIP to extract style embeddings and measure style similarity.
Aesthetic appeal: InternVL2 extracts 40 visual attributes; an aesthetic scorer trained on AVA + BAID outputs an aesthetic score.

Aesthetic evaluation model trained with InternVL-G — Figure A3. Training an aesthetic evaluation model with InternVL-G for multi-dimensional quality scoring.

OmniStyle-1M style and content distribution statistics — Figure A4. Statistics of OmniStyle-1M: style category distribution and stylized-image distribution by content category.

02·BFidelity Data — Destylization-driven (OmniStyle2)

Goal: treat real artistic images as the only ground truth, constructing a high-fidelity style-transfer triplet dataset by reversing the stylization process.

Instead of synthesizing stylized images, we start from real artworks and remove style to recover natural content — the destylized image becomes the input, the original artwork becomes the target.

Data construction pipeline

1) Real artwork collection (Target): 100K+ high-quality artistic images covering 500+ artists.

2) Multi-stage destylization (Destylized): global general destylization → category-wise fine destylization → specialized destylization model.

Figure B1. Multi-stage destylization pipeline: from real-world artwork collection through Stage 1 (global) → Stage 2 (category-wise) → Stage 3 (specialized model), with DeStyleCoT-Filter at each stage.

CoT-driven interpretable quality evaluation

A chain-of-thought reinforced multimodal model evaluates destylized results in an interpretable way, along two dimensions:

Content preservation: region-level object identification + object-wise consistency check.
Style removal: style-attribute identification + attribute-wise difference analysis.

DeStyleCoT-Filter structured reasoning — Figure B2. DeStyleCoT-Filter: structured, interpretable reasoning for content preservation and style removal produces high-quality supervision signals.

04Application Value

The OmniStyle and OmniStyle2 works built on SuperStyle were accepted to CVPR 2025 and ECCV 2026, and highly recognized by reviewers:

Reviewers regard OmniStyle as “solid and impactful work.”
Reviewers regard OmniStyle2 as “a new paradigm for style transfer.”

SuperStyle shifts style transfer toward a supervised learning paradigm, providing a solid data foundation for controllable, high-quality, generalizable style transfer.

Diverse high-quality stylization results on the Eiffel Tower — Figure C1. Diverse, high-quality stylization results — SuperStyle supports strong style-transfer foundation-model training.

Diverse high-quality stylization results on the Golden Gate Bridge — Figure C2. Reference-guided stylization across a wide range of artistic styles.

Roadmap

Domestic adaptation: large-scale style transfer with Jittor (计图) to promote localized models.
Data upgrade: further scale up SuperStyle and strengthen the filtering mechanism.
Continuous maintenance: maintain the project pages & code repos, keeping the open-source releases updated.
Collaboration: engage with universities and industry to improve and apply the dataset.

03Open Source

SuperStyle is fully open-sourced with standardized project pages, a well-structured GitHub repository, and openly downloadable Hugging-Face datasets. It has been downloaded 824+ times since its release on 2025.07.14 (~300 GB total).

Project · BreadthOmniStyle project page
Project · FidelityOmniStyle2 / DeStyle project page
Codegithub.com/StyleX-Research/OmniStyle
Dataset · Breadthhuggingface.co/datasets/StyleXX/OmniStyle-150k
Dataset · Fidelityhuggingface.co/datasets/StyleXX/DeStyle

Hugging Face dataset repository page — Figure D1. Hugging-Face dataset release with detailed docs, code and pretrained models.

The repositories include configuration guides, pretrained-model downloads, usage instructions, and compute-configuration notes. The entire SuperStyle data-generation pipeline is fully automatic and requires no manual intervention, giving it strong scalability for continual model-capability improvement under the supervised paradigm.

SummaryWhy SuperStyle matters

Open & standardized: clean project pages, well-organized GitHub repos, and openly downloadable Hugging-Face datasets.
Advancing the field: shifts style transfer to a supervised paradigm and fuels style-transfer foundation models and generalizable transfer.
Widely used: a high-quality dataset (500K triplets, 300 GB) with 824+ cumulative downloads since 2025.07.14.
Sustainable: a fully-automatic generation pipeline enabling continual capability growth under supervised learning.