Skywork / Unipic3

🌌 UniPic3-Teacher-Model

📖 Introduction

UniPic3-Teacher-Model is the high-quality teacher diffusion model used in the UniPic 3.0 framework. It is trained with full multi-step diffusion sampling and optimized for maximum perceptual quality, semantic consistency, and realism .

This model serves as the teacher backbone for:

• Distribution Matching Distillation (DMD)
• Consistency / trajectory distillation
• Few-step student model training

Rather than being optimized for fast inference, the teacher model prioritizes generation fidelity and stability , providing a strong and reliable supervision signal for downstream distilled models.

🧠 Model Characteristics

• Role : Teacher model (not a distilled student)
• Sampling : Multi-step diffusion (high-fidelity)
• Architecture : Unified UniPic3 Transformer
• Tasks Supported :
  • Single-image editing
  • Multi-image composition (2–6 images)
  • Human–Object Interaction (HOI)
• Resolution : Flexible, within pixel budget constraints
• Training Objective :
  • Flow Matching / Diffusion loss
  • Used as teacher for DMD & consistency training

📊 Benchmarks

This teacher model achieves state-of-the-art performance on:

• Image editing benchmarks
• Multi-image composition benchmarks

It provides high-quality supervision targets for distilled UniPic3 student models.

⚠️ Important Note

This repository hosts the teacher model.
It is not optimized for few-step inference .

If you are looking for:

• ⚡ 4–8 step fast inference
• 🚀 Deployment-friendly distilled models

please refer to the UniPic3-DMD / distilled checkpoints instead.

🧠 Usage (Teacher Model)

1. Clone the Repository

git clone https://github.com/SkyworkAI/UniPic
cd UniPic-3

2. Set Up the Environment

conda create -n unipic python=3.10
conda activate unipic3
pip install -r requirements.txt

3.Batch Inference

transformer_path = "Skywork/Unipic3"

python -m torch.distributed.launch --nproc_per_node=1 --master_port 29501 --use_env \
    qwen_image_edit_fast/batch_inference.py \
    --jsonl_path data/val.jsonl \
    --output_dir work_dirs/output \
    --distributed \
    --num_inference_steps 50 \
    --true_cfg_scale 4.0 \
    --transformer transformer_path \
    --skip_existing

📄 License

This model is released under the MIT License.

Citation

If you use Skywork-UniPic in your research, please cite:

@misc{wang2025skyworkunipicunifiedautoregressive,
      title={Skywork UniPic: Unified Autoregressive Modeling for Visual Understanding and Generation}, 
      author={Peiyu Wang and Yi Peng and Yimeng Gan and Liang Hu and Tianyidan Xie and Xiaokun Wang and Yichen Wei and Chuanxin Tang and Bo Zhu and Changshi Li and Hongyang Wei and Eric Li and Xuchen Song and Yang Liu and Yahui Zhou},
      year={2025},
      eprint={2508.03320},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.03320}, 
}

@misc{wei2025skyworkunipic20building,
      title={Skywork UniPic 2.0: Building Kontext Model with Online RL for Unified Multimodal Model}, 
      author={Hongyang Wei and Baixin Xu and Hongbo Liu and Cyrus Wu and Jie Liu and Yi Peng and Peiyu Wang and Zexiang Liu and Jingwen He and Yidan Xietian and Chuanxin Tang and Zidong Wang and Yichen Wei and Liang Hu and Boyi Jiang and William Li and Ying He and Yang Liu and Xuchen Song and Eric Li and Yahui Zhou},
      year={2025},
      eprint={2509.04548},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.04548}, 
}

@misc{wei2026skyworkunipic30unified,
      title={Skywork UniPic 3.0: Unified Multi-Image Composition via Sequence Modeling}, 
      author={Hongyang Wei and Hongbo Liu and Zidong Wang and Yi Peng and Baixin Xu and Size Wu and Xuying Zhang and Xianglong He and Zexiang Liu and Peiyu Wang and Xuchen Song and Yangguang Li and Yang Liu and Yahui Zhou},
      year={2026},
      eprint={2601.15664},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.15664}, 
}