- Bohgyu Kim, Kyongmo Kim, Nayoung U, DongBin Na
- Konkuk Univ, KAIST Univ, Ajou Univ, POSTECH
Recently, virtual hair styling service using deep learning technology is attracting attention. The core of virtual hair styling technology is ① to change the hairstyle naturally while maintaining the identity of the person, and ② to return the results quickly enough to satisfy the service users. However, existing deep learning-based methods have a problem in that they change not only the hairstyle but also the facial features that indicate the identity of the face, or that inference takes a long time. In this paper, we propose Barbershop++, a method that applies an encoding network to the core components of the existing Barbershop method. Barbershop++ speeds traditional reasoning while maintaining image quality with virtual hair styling. Reduced by 1/3 for improvement. In particular, this paper is expected to be a cornerstone of the deep learning-based virtual hair styling industry/research field in the future in that it can generate natural images even from Korean facial images.
| HairCLIP | Barbershop | Barbershop++ (Ours) |
|---|---|---|
| 11.4 | 864.75 | 278.86 |
- Prepare 2 images (Target, Reference)
- Please put the target image in the form of
Barbershop/unprocessed/input_img.jpg. - Please put the reference image in the form of
Barbershop/unprocessed/ref_img.jpg.
- Please put the target image in the form of
Please download the II2S
and put them in the Barbershop/input/face folder.
Please check the ./Barbershop++-inference.ipynb file in this repository.
After encoding pSp, save .npy file in Barbershop/output/W+ folder
# Align the face of the file in the unprocessed folder, and automatically create an image in the input/face folder
!python align_face.py# inference script example
!python main.py --im_path1 input_img2.png --im_path2 117.png --im_path3 117.png --sign realistic --smooth 1