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:

@article{wang2025skywork,
  title={Skywork unipic: Unified autoregressive modeling for visual understanding and generation},
  author={Wang, Peiyu and Peng, Yi and Gan, Yimeng and Hu, Liang and Xie, Tianyidan and Wang, Xiaokun and Wei, Yichen and Tang, Chuanxin and Zhu, Bo and Li, Changshi and others},
  journal={arXiv preprint arXiv:2508.03320},
  year={2025}
}

@article{wei2025skywork,
  title={Skywork unipic 2.0: Building kontext model with online rl for unified multimodal model},
  author={Wei, Hongyang and Xu, Baixin and Liu, Hongbo and Wu, Cyrus and Liu, Jie and Peng, Yi and Wang, Peiyu and Liu, Zexiang and He, Jingwen and Xietian, Yidan and others},
  journal={arXiv preprint arXiv:2509.04548},
  year={2025}
}

@article{wei2026skywork,
  title={Skywork UniPic 3.0: Unified Multi-Image Composition via Sequence Modeling},
  author={Wei, Hongyang and Liu, Hongbo and Wang, Zidong and Peng, Yi and Xu, Baixin and Wu, Size and Zhang, Xuying and He, Xianglong and Liu, Zexiang and Wang, Peiyu and others},
  journal={arXiv preprint arXiv:2601.15664},
  year={2026}
}