VAREdit

This is the official repository of ICLR 2026 paper: Visual Autoregressive Modeling for Instruction-Guided Image Editing

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

Model Variant	Resolutions	HuggingFace Model	Time (H800)	VRAM (GB)
VAREdit-8B-512	512×512	VAREdit-8B-512	~0.7s	50.41
VAREdit-8B-1024	1024×1024	VAREdit-8B-1024	~1.99s	50.41

🚀 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
pip install flash_attn

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

Parameter	Description	Default
cfg	Classifier-free guidance scale	3.0
tau	Temperature for sampling	0.1
seed	Random seed for reproducibility	-1 (random)

📂 Project Structure

VAREdit/
├── infer.py              # Main inference script
├── train.py              # Main training script
├── trainer.py            # Main trainer 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

Method	Size	EMU-Edit Bal.	PIE-Bench Bal.	Time (A800)
InstructPix2Pix	1.1B	2.923	4.034	3.5s
UltraEdit	7.7B	4.541	5.580	2.6s
OmniGen	3.8B	4.674	3.492	16.5s
AnySD	2.9B	3.129	3.326	3.4s
EditAR	0.8B	3.305	4.707	45.5s
ACE++	16.9B	2.076	2.574	5.7s
ICEdit	17.0B	4.785	4.933	8.4s
VAREdit (256px)	2.2B	5.565	6.684	0.5s
VAREdit (512px)	2.2B	5.662	6.996	0.7s
VAREdit (512px)	8.4B	7.892	8.105	1.2s
VAREdit (1024px)	8.4B	7.379	7.688	3.9s

Note: The released 8B models are trained longer and on more data.

📄 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:

@inproceedings{varedit2026,
  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},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

🙏 Acknowledgments

• Built on the Infinity models

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