Face editing using a regression-based approach in the StyleGAN latent space
Saeid Motiian (Adobe),* Siavash Khodadadeh (University of Central Florida), Shabnam Ghadar (Adobe), Baldo Faieta (Adobe), Ladislau Boloni (University of Central Florida)
The 33rd British Machine Vision Conference
Abstract
Despite significant progress, StyleGAN-based face editing is still limited by undesirable attributes dependencies and artifacts that decrease the quality of generated images. While more, well-annotated training data would likely improve on these problems, collecting such data at scale is very expensive. We propose a face editing architecture that significantly improves the image quality, allows precise specification of individual attributes, and facilitates the introduction of new attributes. We take advantage of recent advances that couple the creation of a latent representation of an image with associated natural language as well as techniques that find linear correlations between the GAN latent space and the attributes of the image, enabling regression models. Our approach deploys carefully chosen regularization approaches that are critical to the integration of these techniques. We demonstrate the ability to edit photorealistic images of faces, originating both from GAN generation and from real images through GAN inversion.Video
Citation
@inproceedings{Motiian_2022_BMVC,
author = {Saeid Motiian and Siavash Khodadadeh and Shabnam Ghadar and Baldo Faieta and Ladislau Boloni},
title = {Face editing using a regression-based approach in the StyleGAN latent space},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0522.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by