AttenCraft: Attention-guided Disentanglement of Multiple Concepts for Text-to-Image Customization

This is the implementation of the paper "AttenCraft: Attention-guided Disentanglement of Multiple Concepts for Text-to-Image Customization". Paper Link

Getting Started

Intsall environment:

conda create --name attencraft --file environment.yml
conda activate attencraft

Training AttenCraft

bash train.sh

Note that the --output_dir flag specifies the output directory where the checkpoints will be saved, and should contain 'wkwv' since this will be used in the inference script.

Inference

Input the chackpoint path, output path, and the text ptompt for the image generation in the inference.py file and run the python script.

Citation

If you find this work helpful, please consider citing the following BibTeX entry:

@article{shentu2024attencraft,
  title={AttenCraft: Attention-guided Disentanglement of Multiple Concepts for Text-to-Image Customization},
  author={Shentu, Junjie and Watson, Matthew and Moubayed, Noura Al},
  journal={arXiv preprint arXiv:2405.17965},
  year={2024}
}