The devil is in the details: Delving into large-pose blind face restoration

⚔️Installation

pip install -r requirements.txt

🧬Pretrained Models And Datasets

The pretrained model weights, test_set and partial tranining set from FFHQ can be found on the BaiduYun.

⚔️Inference

We prepare some examples to show how to do inference.

: Stage I Inference

python inference_face.py  --config ./configs/model/cldm_twoS.yaml  --input input_path --ckpt ./checkpoints/stage1.ckpt --output output_path --device cuda 

: Stage II Inference

python inference_face.py  --config ./configs/model/cldm_twoS_wD.yaml  --input input_path --ckpt ./checkpoints/stage2.ckpt --output output_path   --device cuda 

: Stage III Inference

python inference_face.py  --config ./configs/model/cldm_twoS_AFR.yaml  --use_afr --input input_path  --ckpt ./checkpoints/stage2.ckpt --output output_path   --device cuda 

🌠Train

1. Generate file list of training set and validation set, a file list looks like:

   /path/to/image_1
   /path/to/image_2
   /path/to/image_3

   We prepare a script for you to generate the file list: make_file_list.

   python make_file_list.py --img_folder ./data_folder --save_path ./test.list 
   

2. Fill in the train and val configuration file with yout file list script: face*.yaml.

3. training stage I

   HF_ENDPOINT=https://hf-mirror.com python train.py --config ./configs/train_stage1.yaml

4. training stage II

   HF_ENDPOINT=https://hf-mirror.com python train.py --config ./configs/train_stage2.yaml

5. generate latent code to reduce computation cost:

   python inference_face.py  --config ./configs/model/cldm_twoS_wD.yaml --generate_latent  --ckpt ./checkpoints/stage2.ckpt --output ./results/test_generate  --device cuda 

   This code will generate the latent code to ' ./results/test_generate ', please use make_file_list to generate file list and replace the path in stage3*.yaml

6. training stage III

   HF_ENDPOINT=https://hf-mirror.com python train.py --config ./configs/train_stage3.yaml

Acknowledgement

This project is based on ControlNet, BasicSR, StableSR, DiffBIR, GFPGAN, DifFace. Thanks for their awesome work.