A discovery of the novel artistic technique: Databrushing while researching with Generative Adversarial Networks (GANS).

Sometimes the best discoveries lay in pivots, progression, and abstraction.


Delving deeply into the applied world of applied Artificial Intelligence concepts (Kudos to CS334) we were tasked to contribute to an existing branch of narrow AI.

My primary interests being in Generative Adversarial Networks (GAN’s), i had decided to launch an incremental analysis of human error on the effects of StyleGan model training. Human error in this scope being the ‘mis-categorization’ of an image into the wrong classifier.

Midway through the project, a dazzling visual discovery arose. A combination of a strict texture + form dataset yielded fusions of incredibly detailed and creative images. With the warm encouragement of my professor, this discovery was explored further to reveal the novel technique: Databrushing. 

As an artistic technique, Databrushing can be best described as a method of collecting visual data much in the way an artist mixes paints to create a unique color. Once this color is mixed, the artist uses the brush, StyleGAN2, as the instrument for creating art from their tailored palette. The paint, in this analogy, is the image dataset supplied for model training. By ‘mixing” images types such as “crystal textures” and “roses” we were able to create latent spacewalks and images worthy of an avant-garde art exhibit.

50:50 Crystal | Rose
Latent Space Walk 50:50 Crystal | Rose
50:50 Crystal | Rose
75:25 Crystal | Rose
75:25 Crystal | Rose

Research Effective

With DataBrushing, we can take a closer look into how progressive growing GAN’s parse features/textures. This method let’s us peek a bit closer at the otherwise 'black box’ theory of generative networks. Although further research should be conducted, preliminary results from Databrushing also show support for the texture bias hypothesis of Geirhos 2018 et al. in generative models.

To read more, check back here June 2020 for the full medium article and Research Brief

US Dept. of Homeland Security
Patrol Agent