This is the official implementation of paper DeepFacePencil: Creating Face Images from Freehand Sketches. arXiv, Project
Also see another project: Lines2Face project
PyTorch 1.6, Python 3.7, NumPy, scipy, PIL, tqdm
We use CelebAMask-HQ to obtain synthesized (boundary) sketches. Deformed sketches are generated by vectorizing synthesized sketches using AutoTrace and adding random offsets to endpoints and control points of vectorized strokes.
The test sketches should be put in folder ./datasets/CelebAMask/test_A. The pretrained model should be put in folder ./checkpoints/pretrained. Example script of testing can be found in ./test_scripts.sh. The results are supposed to be in ./results. The deform.pth file is uploaded to Google Drive.
@inproceedings{DeepFacePencil,
author = {Li, Yuhang and Chen, Xuejin and Yang, Binxin and Chen, Zihan and Cheng, Zhihua and Zha, Zheng-Jun},
title = {DeepFacePencil: Creating Face Images from Freehand Sketches},
booktitle = {Proceedings of the 28th ACM International Conference on Multimedia},
series = {MM '20},
year = {2020},
isbn = {978-1-4503-7988-5/20/10},
location = {Seattle, WA, USA},
pages = {},
numpages = {9},
url = {http://doi.acm.org/10.1145/3394171.3413684},
doi = {10.1145/3394171.3413684},
acmid = {3413684},
publisher = {ACM},
address = {New York, NY, USA},
keywords = {Sketch-based synthesis, face image generation, spatial attention, dual generator, conditional generative adversarial networks},
}
This code borrows heavily from pix2pixHD and pytorch-CycleGAN-and-pix2pix.