RestorerID: Towards Tuning-Free Face Restoration with ID Preservation

Arxiv

Install

# Create a conda environment and activate it
conda env create --file environments.yaml
conda activate RestorerID

# Install xformers
conda install xformers -c xformers/label/dev

# Install taming-transfomers
pip install -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
pip install -e .

Datasets

FFHQ
VGGFace2
Celeb-Ref

Inference

# You can download the pretrained model from hugging-face and put it in ckpt path: ckpt/RestorerIDFull.ckpt
[RestorerID-huggingface](https://huggingface.co/YingJiacheng/RestorerID/)

# run
bash inference.sh
or
CUDA_VISIBLE_DEVICES=0 python scripts/Inference.py --LQpath TestSamples/1/lq1.png  --Refpath TestSamples/1/ref1.png  --Outputpath Results/1/

Train

# First train base model
# download sdv15 pretrained model (runwayml/v1-5-pruned.ckpt) from huggingface, put into the ckpt path as: ckpt/v1-5-pruned.ckpt
# prepare your datasets

CUDA_VISIBLE_DEVICES=0,1 python train_basemodel.py --train --base configs/v15/v15-BaseModel.yaml  --name v15_basemodel --scale_lr False

# Then train RestorerID
# rename your trained base model and put to ckpt path as: ckpt/basemodel.ckpt
# download ID model ip-adapter-faceid-plus_sd15.bin from [IPAdapter-huggingface](https://huggingface.co/h94/IP-Adapter-FaceID/tree/main), put it into the ckpt path as: ckpt/ip-adapter-faceid-plus_sd15.bin
# prepare your datasets
# combine basemodel with ID model

python Combineckpt.py
CUDA_VISIBLE_DEVICES=0,1 python train_RestorerID.py --train --base configs/v15/v15-RestorerID.yaml  --name RestorerID --scale_lr False

License

This project is released under the Apache 2.0 license.

Acknowledgement

This work is mainly based on StableSR, IPAdapter, we thank the authors for the contribution.