Tucson Sculpture Festival 2012

Monday, 24 August 2020

Looking for Faces -- StyleGAN2 Default Settings in Runway ML

I revisited my "model" in Runway ML ("trained" in StyleGAN2), and cherry-picked a bunch of the AI generated drawings in my style with better FACES (without changing the default settings).  NOTE:  I did not sketch any of these figures, the computer generated these fake figure drawings:


I did not draw this figure and face,
the computer did


Then I further altered the results by applying a color and style to them in Deep Dream Generator, and then doing a little post-editing afterwards in Photoshop Elements.


I colored and enhanced this fake drawing
with another AI program online


The interface is a little awkward, but I did figure out how to generate new variations of images in my drawing style, and download several of those that I liked.  Even with the larger thumbnails, it's still unclear what I'm downloading, until I open them up in another program. 


Default Settings --
Truncation: point 8
Sampling Distance: point 5





























TWIST: In Deep Dream Generator I applied the style of one of my abstract drawings (see below), to the computer generated fake drawing above, to get the result (two images below).






Overall I am very impressed with the fake figure drawings that the computer generates.  The program often preserves the weight and gesture of my style, which are the priorities in my drawing approach.  The fake drawings still obey gravity, more times than not.

The fake faces suffer.  However, more often than not, I also mess up the face in my original figure drawings.  The computer program is too smart, and continues to mess up the faces more than the bodies, perpetuating my weaknesses.  It's amazing that the computer program knows the difference between the portrait and the body to begin with.


No comments:

Post a Comment

Note: only a member of this blog may post a comment.