Ye Wang¹, Ruiqi Liu¹, Jiang Lin², Fei Liu³, Zili Yi², Yilin Wang^4,*, Rui Ma^1,5,*

¹School of Artificial Intelligence, Jilin University   
²School of Intelligence Science and Technology, Nanjing University   
³ByteDance ⁴Adobe   
⁵Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MOE, China
^*Corresponding authors

• [2025.07.23] OmniStyle-150K dataset is now available!
• [2025.07.11] Code and model weights for OmniStyle are now available!
• [2025.07.05] Released the project page.

• ✅ Release Model weights and inference code for OmniStyle.
• ✅ Release OmniStyle-150K: The filtered high-quality subset used for training.

🤖 OmniStyle is the first end-to-end style transfer framework based on the Diffusion Transformer (DiT) architecture, achieving high-quality 1K-resolution stylization by leveraging the large-scale, filtered OmniStyle-1M dataset. It supports both instruction- and image-guided stylization, enabling efficient and versatile style transfer across diverse styles.

🗂️ OmniStyle-1M is the first million-scale paired style transfer dataset, comprising over one million triplets of content, style, and stylized images across 1,000 diverse style categories. It provides strong supervision for learning controllable and generalizable style transfer models.

🧪 OmniStyle-150K is a high-quality subset of OmniStyle-1M, specifically filtered to train the OmniStyle model.

We recommend creating a clean conda environment:

conda create -n omnistyle python=3.10 
conda activate omnistyle
# Install dependencies
pip install -r requirements.txt

You can download the pretrained OmniStyle model from Hugging Face:

👉 https://huggingface.co/StyleXX/OmniStyle

After downloading, please place the .safetensors checkpoint file into the ./ckpts/ directory:

In addition, you should download relevant model weights from FLUX-Dev:

👉 https://github.com/XLabs-AI/x-flux

After downloading all weights, you need to specify the correct checkpoint paths in test.sh:

We have provided example style and content images in the test/ folder.

To run image-guided stylization, simply execute:

CUDA_VISIBLE_DEVICES=0 python inference_img_guided.py

The generated results will be saved in the output/ folder.

For instruction-guided stylization, just run:

CUDA_VISIBLE_DEVICES=0 python inference_instruction_guided.py

As with image-guided transfer, the results will be saved in the output/ folder.

OmniStyle supports high-resolution (1k) image stylization. Below are the typical GPU memory usages during inference:

📌 Note: For stable inference, please ensure at least 48 GB available GPU memory.

💡 Recommendation: OmniStyle is optimized for 1024×1024 resolution. We recommend using this resolution during inference to achieve the best stylization quality.

Our code is built with reference to the following excellent projects. We sincerely thank the authors for their open-source contributions:

• x-flux
• UNO

Their work greatly inspired and supported the development of OmniStyle.