PEA-Diffusion (ECCV 2024)

The official code for the paper PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation.

Checkpoint

Checkpoint path https://huggingface.co/OPPOer

Introduction

We are inspired to propose a simple plug-and-play language transfer method based on knowledge distillation. All we need to do is train a lightweight MLP-like parameter-efficient adapter (PEA) with only 6M parameters under teacher knowledge distillation along with a small parallel data corpus. We are surprised to find that freezing the parameters of UNet can still achieve remarkable performance on the language-specific prompt evaluation set, demonstrating that PEA can stimulate the potential generation ability of the original UNet. Additionally, it closely approaches the performance of the English text-to-image model on a general prompt evaluation set. Furthermore, our adapter can be used as a plugin to achieve significant results in downstream tasks in cross-lingual text-to-image generation.

Requirements

A suitable conda environment named PEA-Diffusion can be created and activated with:

conda create -n PEA-Diffusion   
source activate PEA-Diffusion   
pip install -r requirements.txt

Data Prepare

The English data we trained directly used LAION, and the Chinese data came from WuKong, LAION_ZH. Our training data is webdataset format. If only multilingual-CLIP training is required, then only need English image-text pairs. If training an PEA for a language-specific and aiming to generate images that are culturally relevant to that language, parallel corpora are necessary. We suggest download the data or translating the English prompts into the specific language. For more details, please refer to our paper.

Training based SDXL

bash train_sdxl_zh.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_sdxl_zh.sh.

The training code includes a large number of model paths that need to be downloaded by yourself. For detailed download paths, please see Appendix 6.2 of the paper.

train_sdxl.py can run T21 in four other languages: Italian, Russian, Korean, and Japanese. Similarly, you need to download the corresponding clip model and put it in the corresponding path.

Inference

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

python tests/test_sdxl_zh.py

For more downstream test scripts, please view the tests directory

Downstream Performance

The PEA module can be easily applied to a variety of downstream tasks with plug-and-play,The figure below shows seven common downstream tasks.

TODOs

• [x] Release inference code
• [x] Release training code
• [ ] Release PEA checkpoint
• [ ] Release demo

Acknowledgements

We borrow some code from TorchData

Citation

@misc{ma2023peadiffusion,
      title={PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation}, 
      author={Jian Ma and Chen Chen and Qingsong Xie and Haonan Lu},
      year={2023},
      eprint={2311.17086},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}