Project Page | Paper | Demo | Model | Introduction Video

English | 简体中文

📋 Overview

IC-Custom is designed for diverse image customization scenarios, including:

• Position-aware: Input a reference image, target background, and specify the customization location (via segmentation or drawing)

  • Examples: Product placement, virtual try-on

• Position-free: Input a reference image and a target description to generate a new image with the reference image's ID

  • Examples: IP customization and creation

---

✨ Note

• This is the first release of IC-Custom (0x1561) 🎉.
• The model is currently stronger at masked customization for categories such as bags 👜, perfumes 💎, clothes 👗, and rigid objects.
• Training data ensures quality (≥800px ✅) but is still limited in scale and diversity. Face-related data and style transfer are not yet included.
• The position-free ability (IP customization without masks), along with other capabilities such as face customization, is still being improved and will be further refined based on community feedback 🔄.
• We are also exploring acceleration techniques (quantization, distillation) and finer-grained customization (e.g., flexible control of product viewpoints 👀).

Community Support

• ComfyUI: ComfyUI_RH_ICCustom

• RunningHub:

  • IC Custom角色+场景迁移 V1
  • IC Custom万物迁移双图参考 V1
  • IC Custom换装迁移 V1
  • IC Custom万物迁移单图参考 V1

• Demonstration video:

  • T8star-Aix ComfyUI demo (Bilibili)
  • Comfyui官方教程 (Bilibili)

Acknowledgements: ComfyUI deployment support by HM-RunningHub; RunningHub workflows by T8star-Aix.

---

📑 Table of Contents

• 📋 Overview
• ✨ Note
• 📑 Table of Contents
• 🚀 Environment Requirements
• 🔧 Installation
• 📦 Model Checkpoints
• 💻 Running Scripts
  • App (Gradio Interface)
  • Inference
• 📅 Update Logs
• 📝 Citation
• 💖 Acknowledgements
• Limitation
• 📄 License
• 📬 Contact
• 🌟 Star History

---

🚀 Environment Requirements

IC-Custom has been implemented and tested on:

• CUDA 12.4
• PyTorch 2.6.0
• Python 3.10.16

🔧 Installation

1. Clone the repository

   git clone https://github.com/TencentARC/IC-Custom.git
   cd IC-Custom

2. Set up Python environment

   conda create -n ic-custom python=3.10 -y
   conda activate ic-custom
   pip install -e .
   pip install -r requirements.txt

3. Custom CUDA versions (optional) If you require a different CUDA version, you can ignore the torch-related packages listed in requirements.txt. Instead, please install PyTorch and xformers that are compatible with your CUDA version by following the instructions on the official PyTorch website.

   Example for CUDA 12.4:

   pip3 install xformers torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --index-url https://download.pytorch.org/whl/cu124

4. Flash Attention 2 (optional) If your device supports Flash Attention 2, you can optionally install flash-attn. We use flash-attn==2.7.3. Find compatible versions at flash-attention releases.

📦 Model Checkpoints

You'll need to obtain the model checkpoints before running the inference or app.

• Option 1 — Automatic download: checkpoints are fetched when running scripts (requires valid HF_TOKEN).
• Option 2 — Manual download: use the helper script and then set paths in configs.

sh scripts/inference/download_models.sh $HF_TOKEN

Expected directory structure (example):

|-- models
    |-- clip-vit-large-patch14
    |-- ic-custom
    |   |-- dit_boundary_embeddings_0x1561.safetensors
    |   |-- dit_lora_0x1561.safetensors
    |   |-- dit_task_register_embeddings_0x1561.safetensors
    |   |-- dit_txt_img_in_0x1561.safetensors
    |   ...
    |-- siglip-so400m-patch14-384
    |-- t5-v1_1-xxl
    |-- ae.safetensors
    |-- flux1-fill-dev.safetensors
    |-- flux1-redux-dev.safetensors

After manual download, edit the YAMLs in configs/** (e.g., inference.yaml) to point to your local model paths.

• A compact per-model list with links is available in MODEL_CARD.md.

💻 Running Scripts

App (Gradio Interface)

When running the app, all required models are automatically downloaded. Optional models are fetched only when explicitly enabled.

sh src/app/run_app.sh $HF_TOKEN $HF_CACHE_DIR

For required vs optional models and configuration, see APP.md.

Inference

Run the inference script with your Hugging Face token:

sh scripts/inference/inference.sh $HF_TOKEN $HF_CACHE_DIR

Parameters:

• $HF_TOKEN: Your Hugging Face access token (required for automatic model download, optional if model paths are specified in configs/**/*.yaml)
• $HF_CACHE_DIR (optional): Custom cache directory for downloaded models (default: "~/.cache/huggingface/hub")

📅 Update Logs

• 2025/07/03 - Released paper, webpage.
• 2025/08/26 - Released Checkpoint v0x1561, APP, and Inference Code.
• TBD - Test and Training Code.
• TBD - A stronger model version.

📝 Citation

@article{li2025ic,
  title={IC-Custom: Diverse Image Customization via In-Context Learning},
  author={Li, Yaowei and Li, Xiaoyu and Zhang, Zhaoyang and Bian, Yuxuan and Liu, Gan and Li, Xinyuan and Xu, Jiale and Hu, Wenbo and Liu, Yating and Li, Lingen and others},
  journal={arXiv preprint arXiv:2507.01926},
  year={2025}
}

💖 Acknowledgements

We gratefully acknowledge the use of code from:

• flux
• x-flux

We also thank Hugging Face for providing professional model hosting and Spaces for deployment.

Limitation

Current limitations are mainly inference speed and the need for more flexible instruction following. We plan to improve these aspects in future updates. If you have any feedback, please feel free to contact us.

📄 License

We are pleased to support the open source community. For complete license details, see LICENSE and NOTICE.

📬 Contact

For any questions, feel free to email.

🌟 Star History