Zødiac

Code & 3d Art

StyleGANs Experiments

Python - Pytorch - StyleGAN 2

November 2021

I started from the paper Audio-reactive Latent Interpolations with StyleGAN from Hans Brouwer and some pretrained models:

The latent movements are based on:

The audio input you can hear (make sure to unmute the videos)
A looped slerp interpolation, which is added on top of it to ensure a variety of latent space positions

I then started non scientific experiments by swapping hidden layers between the pretrained models to see how it would affect the result based on the depth of the layer swapped, and the number of layers swapped.
Here are some cool results:

In the 1st video I swapped half of the layers and you can clearly see that:

The FFHQ model dictates the high level features of the images (mostly shapes)
The Abstract art model is mostly responsible of the textures used

In the 2nd and 3rd videos I swapped only one layer, but changed the depth of the swapped layer. which results in:

2nd one: the Floor plans model dictates many high level features, and FFHQ makes most of the textures
3rd one: the opposite

For the following videos I'm using the Floor plans and Abstract art models.

In the 1st one the effect of the sound on the latent movements is extreme, causing it to go out of distribution (out of the domain learnt via the training data) which results in a nice abstract piece

Zødiac

Code & 3d Art

3D

Code

Contact

StyleGANs Experiments

Python - Pytorch - StyleGAN 2

November 2021