VAREdit: Visual Autoregressive Modeling for Instruction-Guided Image Editing

📄 Paper

VAREdit is an advanced image editing model built on the Infinity models, designed for high-quality instruction-based image editing.

Try our online demos: 🤗VAREdit-8B-1024 and 🤗VAREdit-8B-512.

🌟 Key Features

• Strong Instruction Follow: Follows instructions more accurately due to the autoregressive nature of the model.
• Efficient Inference: Optimized for fast generation with less than 1 seconds for 8B model.
• Flexible Resolution: Supports 512×512 and 1024×1024 image resolutions

📊 Model Variants

🚀 Quick Start

Prerequisites

Before starting, ensure you have:

• Python 3.8+
• CUDA-compatible GPU with sufficient VRAM (8GB+ for 2B model, 24GB+ for 8B model)
• Required dependencies installed

Installation

1. Clone the repository

git clone https://github.com/HiDream-ai/VAREdit.git
cd VAREdit

2. Install dependencies

pip install -r requirements.txt

3. Download model checkpoints

Download the VAREdit model checkpoints:

# Download from HuggingFace
git lfs install
git clone https://huggingface.co/HiDream-ai/VAREdit

Basic Usage

from infer import load_model, generate_image

model_components = load_model(
    pretrain_root="HiDream-ai/VAREdit",
    model_path="HiDream-ai/VAREdit/8B-1024.pth",
    model_size="8B",
    image_size=1024
)

# Generate edited image
edited_image = generate_image(
    model_components,
    src_img_path="assets/test.jpg",
    instruction="Add glasses to this girl and change hair color to red",
    cfg=3.0,  # Classifier-free guidance scale
    tau=0.1,  # Temperature parameter
    seed=42  # Optional random seed
)

📝 Detailed Configuration

Model Sampling Parameters

📂 Project Structure

VAREdit/
├── infer.py              # Main inference script
├── infinity/             # Core model implementations
│   ├── models/          # Model architectures
│   ├── dataset/         # Data processing utilities
│   └── utils/           # Helper functions
├── tools/               # Additional tools and scripts
│   └── run_infinity.py  # Model execution utilities
├── assets/              # Demo images and resources
└── README.md           # This file

📊 Performance Benchmarks

Note: The released 8B models are trained longer and on more data, so the performances are better than that in the paper.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📚 Citation

If you use VAREdit in your research, please cite:

@article{varedit2025,
  title={Visual Autoregressive Modeling for Instruction-Guided Image Editing},
  author={Mao, Qingyang and Cai, Qi and Li, Yehao and Pan, Yingwei and Cheng, Mingyue and Yao, Ting and Liu, Qi and Mei, Tao},
  journal={arXiv preprint},
  year={2025}
}

🙏 Acknowledgments

• Built on the Infinity models

Note: This project is under active development. Features and code may change.