Exploring the GANformer for Face Generation: Investigating the segmentation and smile augmentation potential

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

532 Downloads (Pure)

Abstract

Advancing the research in face applications is
limited by proprietary databases and increasing data pro-
tection regulations, synthetically generated databases may
provide a solution. In this work the GANformer, a hybrid
generative image model, is explored for this application.
While only trained for unconditioned face generation like
many other models, this works shows the potential of two
use cases. First, the unique implementation of the attention
is examined for the application of segmentation. Second,
real labeled faces are reconstructed in latent space to find
latent directions describing disentangled attributes. This
concept is brought in practice by augmenting neutral to
smiling faces, but could be applied on other expressions
and attributes as well. This work can be use as basis as
it opens up two directions for further research.
Original languageEnglish
Title of host publicationProceedings of the 2022 Symposium on Information Theory and Signal Processing in the Benelux
Pages38
Number of pages8
Publication statusPublished - 1 Jun 2022
Event42nd WIC Symposium on Information Theory and Signal Processing in the Benelux, SITB 2022 - Université catholique de Louvain, Louvain-la Neuve, Belgium
Duration: 1 Jun 20222 Jun 2022
Conference number: 42
https://sites.google.com/view/sitb2022/home

Conference

Conference42nd WIC Symposium on Information Theory and Signal Processing in the Benelux, SITB 2022
Abbreviated titleSITB 2022
Country/TerritoryBelgium
CityLouvain-la Neuve
Period1/06/222/06/22
Internet address

Fingerprint

Dive into the research topics of 'Exploring the GANformer for Face Generation: Investigating the segmentation and smile augmentation potential'. Together they form a unique fingerprint.

Cite this