FreeGraftor

FreeGraftor: Training-Free Cross-Image Feature Grafting for Subject-Driven Text-to-Image Generation

Zebin Yao,   Lei Ren,   Huixing Jiang,   Wei Chen,   Xiaojie Wang,   Ruifan Li,   Fangxiang Feng

🛠️ Installation

git clone https://github.com/Nihukat/FreeGraftor.git
cd FreeGraftor
conda create -n freegraftor python==3.10.9
conda activate freegraftor
pip install -r requirements.txt

📝 Preparation

1. Download SAM checkpoints.

aria2c -x 4 -j 5 https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth 

2. Download FLUX.1 and Grounding DINO from huggingface.

FLUX.1:

huggingface-cli download --token hf_your_huggingface_token --resume-download black-forest-labs/FLUX.1-dev --local-dir your/directory/for/FLUX.1-dev

Grounding DINO:

huggingface-cli download --resume-download IDEA-Research/grounding-dino-tiny --local-dir your/directory/for/grounding-dino-tiny

3. Setup model paths

Set the paths to the downloaded models in your environment variables:

export FLUX_DEV=your/directory/for/FLUX.1-dev
export GROUNDING_DINO=your/directory/for/grounding-dino-tiny
export SAM=./sam_vit_h_4b8939.pth

You can also uncomment and modify the paths in the src/generate.py and src/app.py files:

os.environ['FLUX_DEV'] = 'your/directory/for/FLUX.1-dev'
os.environ['GROUNDING_DINO'] = 'your/directory/for/grounding-dino-tiny'
os.environ['SAM'] = './sam_vit_h_4b8939.pth'

🚀 Usage

Basic Usage

Single-concept examples :

python src/generate.py \
--gen_prompt "A can is placed on a desk, next to a laptop." \
--class_names "beer can" \
--image_paths "examples/can_00.jpg"\
--output_dir "inference_results" \
--start_seed 0 \
--num_images 2 \

You can also use a configuration file to pass parameters of the concept:

python src/generate.py \
--gen_prompt "A can is placed on a desk, next to a laptop." \
--concept_config_path "configs/can.json"
--output_dir "inference_results" \
--start_seed 0 \
--num_images 2 \

Multi-subject examples :

python src/generate.py \
--gen_prompt "A cat and a dog in Times Square, zoom in, close-up. " \
--class_names "cat" "dog" \
--image_paths "examples/cat_01.jpg" "examples/dog_02.jpg" \
--output_dir "inference_results" \
--start_seed 0 \
--num_images 2 \

You can also use a configuration file to pass parameters of the concepts:

python src/generate.py \
--gen_prompt "A cat and a dog in Times Square, zoom in, close-up. " \
--concept_config_path "configs/cat_dog.json"
--output_dir "inference_results" \
--start_seed 0 \
--num_images 2 \

Gradio Demo

python src/app.py

✅ To-Do List

• More Usage and Applications
• Support for More Models
• Visualization Results
• Gradio Demo
• Source Code

💖 Acknowledgements

We sincerely thank FireFlow, RF-Solver and FLUX for their well-structured codebases. The support and contributions of the open-source community have been invaluable, and without their efforts, completing our work so efficiently would not have been possible.

📚 Citation

If you find this code useful for your research, please consider citing:

@article{yao2025freegraftor,
  title={FreeGraftor: Training-Free Cross-Image Feature Grafting for Subject-Driven Text-to-Image Generation},
  author={Yao, Zebin and Ren, Lei and Jiang, Huixing and Wei, Chen and Wang, Xiaojie and Li, Ruifan and Feng, Fangxiang},
  journal={arXiv preprint arXiv:2504.15958},
  year={2025}
}