1073521013/subject-diffusion

Subject-Diffusion:Open Domain Personalized Text-to-Image Generation without Test-time Fine-tuning

Subject-Diffusion

[Project Page] [Paper]

Requirements

A suitable conda environment named subject-diffusion can be created and activated with:

conda env create -f environment.yaml
conda activate subject-diffusion

Data Prepare

First, you need install GroundingDINO. Then run:

python data_process.py tar_path tar_index_begin tar_index_end output_path

The first parameter represents the data path of webdataset image text pair. The original data can be downloaded by img2dataset command; The last two parameters represent the beginning and end of the index for webdataset data

Training

bash train.sh 0 8

The first parameter represents the global rank of the current process, used for inter process communication. The host with rank=0 is the master node. and the second parameter is the world size. Please review the detailed parameters of model training with train_en.sh script

Inference

We provide a script to generate images using pretrained checkpoints. run

python test.py

TODOs

• Release inference code
• Release training code
• Release data preparation code
• Release demo
• Release training data

Acknowledgements

This repository is built on the code of diffusers library. Additionally, we borrow some code from GLIGEN, FastComposer and GlyphDraw.