Building a GAN Model to provide "faces" from a limited dataset

Discussions about research, Faceswapping and things that don't fit in the other categories here.


Post Reply
User avatar
ianstephens
Posts: 117
Joined: Sun Feb 14, 2021 7:20 pm
Has thanked: 12 times
Been thanked: 15 times

Building a GAN Model to provide "faces" from a limited dataset

Post by ianstephens »

So I had a thought about the possibility of building and training a GAN model for generating more "faces" for a limited "face" dataset.

The model would learn on the limited dataset and also the landmarks of the faces. Then, we would ask the model to provide different faces based on different landmark configurations/shapes.

Perhaps an additional large dataset of a well-known face that has many different landmark configurations would also be beneficial to be pre-trained on and frozen or something similar.

This is all new to me but has anyone else had any similar ideas or even implemented such a thing? How feasible would it actually be - and of course - is it really worth the effort?

User avatar
torzdf
Posts: 2621
Joined: Fri Jul 12, 2019 12:53 am
Answers: 155
Has thanked: 128 times
Been thanked: 610 times

Re: Building a GAN Model to provide "faces" from a limited dataset

Post by torzdf »

On the former, although I have not tested myself, the general consensus is that feeding FS with ML generated faces tends to not work out so well (there are artefacts in generated faces that we may not see, but the model definitely does). That being said, this should not dissuade you from experimenting and reporting back :)

The 2nd, yes... absolutely. I do similar. On projects where timings are tight for data collection, I will often 'pretrain' a model on known facesets I have, and then switch out the sources when the real data arrives.

My word is final

Post Reply