Paper | Supplementary | Demo | ModelScope

Tao Yang¹, Peiran Ren¹, Xuansong Xie¹, Lei Zhang^1,2
1DAMO Academy, Alibaba Group, Hangzhou, China
²Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China

(2023-02-15) GPEN-BFR-1024 and GPEN-BFR-2048 are now publicly available. Please download them via [ModelScope2].

(2023-02-15) We provide online demos via [ModelScope1] and [ModelScope2].

(2022-05-16) Add x1 sr model. Add --tile_size to avoid OOM.

(2022-03-15) Add x4 sr model. Try --sr_scale.

(2022-03-09) Add GPEN-BFR-2048 for selfies. I have to take it down due to commercial issues. Sorry about that.

(2021-12-29) Add online demos . Many thanks to CJWBW and AK391.

(2021-12-16) Release a simplified training code of GPEN. It differs from our implementation in the paper, but could achieve comparable performance. We strongly recommend to change the degradation model.

(2021-12-09) Add face parsing to better paste restored faces back.

(2021-12-09) GPEN can run on CPU now by simply discarding --use_cuda.

(2021-12-01) GPEN can now work on a Windows machine without compiling cuda codes. Please check it out. Thanks to Animadversio. Alternatively, you can try GPEN-Windows. Many thanks to Cioscos.

(2021-10-22) GPEN can now work with SR methods. A SR model trained by myself is provided. Replace it with your own model if necessary.

(2021-10-11) The Colab demo for GPEN is available now .

• Install modelscope:

pip install "modelscope[cv]" -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html

• Run the following codes:

import cv2
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from modelscope.outputs import OutputKeys

portrait_enhancement = pipeline(Tasks.image_portrait_enhancement, model='damo/cv_gpen_image-portrait-enhancement-hires')
result = portrait_enhancement('https://modelscope.oss-cn-beijing.aliyuncs.com/test/images/marilyn_monroe_4.jpg')
cv2.imwrite('result.png', result[OutputKeys.OUTPUT_IMG])

It will automatically download the GPEN models. You can find the model in the local path ~/.cache/modelscope/hub/damo. Please note pytorch_model.pt, pytorch_model-2048.pt are respectively the 1024 and 2048 versions.

• Clone this repository:

git clone https://github.com/yangxy/GPEN.git
cd GPEN

• Download RetinaFace model and our pre-trained model (not our best model due to commercial issues) and put them into weights/.

  RetinaFace-R50 | ParseNet-latest | model_ir_se50 | GPEN-BFR-512 | GPEN-BFR-512-D | GPEN-BFR-256 | GPEN-BFR-256-D | GPEN-Colorization-1024 | GPEN-Inpainting-1024 | GPEN-Seg2face-512 | realesrnet_x1 | realesrnet_x2 | realesrnet_x4

• Restore face images:

python demo.py --task FaceEnhancement --model GPEN-BFR-512 --in_size 512 --channel_multiplier 2 --narrow 1 --use_sr --sr_scale 4 --use_cuda --save_face --indir examples/imgs --outdir examples/outs-bfr

• Colorize faces:

python demo.py --task FaceColorization --model GPEN-Colorization-1024 --in_size 1024 --use_cuda --indir examples/grays --outdir examples/outs-colorization

• Complete faces:

python demo.py --task FaceInpainting --model GPEN-Inpainting-1024 --in_size 1024 --use_cuda --indir examples/ffhq-10 --outdir examples/outs-inpainting

• Synthesize faces:

python demo.py --task Segmentation2Face --model GPEN-Seg2face-512 --in_size 512 --use_cuda --indir examples/segs --outdir examples/outs-seg2face

• Train GPEN for BFR with 4 GPUs:

CUDA_VISIBLE_DEVICES='0,1,2,3' python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 train_simple.py --size 1024 --channel_multiplier 2 --narrow 1 --ckpt weights --sample results --batch 2 --path your_path_of_croped+aligned_hq_faces (e.g., FFHQ)

When testing your own model, set --key g_ema.

Please check out run.sh for more details.

If our work is useful for your research, please consider citing:

@inproceedings{Yang2021GPEN,
    title={GAN Prior Embedded Network for Blind Face Restoration in the Wild},
    author={Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang},
    booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2021}
}

© Alibaba, 2021. For academic and non-commercial use only.

We borrow some codes from Pytorch_Retinaface, stylegan2-pytorch, Real-ESRGAN, and GFPGAN.

If you have any questions or suggestions about this paper, feel free to reach me at yangtao9009@gmail.com.