Thursday 31 December 2020

Artificial Intelligence creating Fake Bad Krrrl Drawings

I trained my drawings with the StyleGAN2 program in Runway ML in order to create fake Krrrl drawings.  The AI (artificial intelligence) program created thousands of drawings in my style, sketches that I did not draw.




1st Model: "Book_model_final.pkl" (364 MB
3rd Model: "Long_Krrrl_Drawings__final.pkl" (364 MB



1st MODEL

I uploaded all the 201 high resolution drawings from my book into the StyleGAN2 program and made the 1st "model" that then produced infinite fake drawings in my style.  I "trained" on the model of bird drawings.

(1024 x 1024 images)






I then applied color and style to the Fake Drawings using another AI program -- Deep Dream Generator :



While I was impressed with the Fake Drawings that the computer generated, it often obliterated the faces.  So I went back into the model and cherry picked the Fake Drawings with the best faces:



I played with changing the settings of the sampling distance -- Truncation .8, Sampling 1.  The Fake Drawings looked more coherent.



I changed to the highest settings -- Truncation 1, Sampling 1 -- and was more pleased with the resulting Fake Drawings.  The lowest settings -- Truncation 0, Sampling 0 -- did not work, just giving me a single Fake Drawing.



How far could I push my Fake Drawings?





Later I uploaded 1456 short pose drawings from this blog to make the 2nd"model" that produced a little different fake drawings in my style.  The larger data set is supposed to give better results. I "trained" this model on the 1st Model I made.


Lastly I uploaded 348 long pose drawings to make the 3rd "model"   I "trained" this model on my 2nd Model.  I think this model generated the best Fake Drawings.

(1024 x 1024 images)


 





Note:  While I made this 3rd Model public on Runway ML, it has only been run once, making it probably the least popular model of all time.

ARTIFICIAL IGNORANCE


I wanted to subvert the machine learning system in order to create something entirely different.



To create something unexpected, I started with the same process as with the 1st Model -- training the 201 drawings from my book on a model of birds. However this time I arrested the training process at only 1000 steps, so that the drawings were half bird/half Krrrl.




*****

RUNWAY ML TUTORIALS


I watched the Runway ML tutorials by Derrick Shultz and Lia Coleman.


UPDATE (January 22, 2022):  I posted the 3rd model to Justin Pinkney's "awesome pretrained stylegan2" collection on Github, and Mark Hanslip from the UK used it to make an audio-visual presentation  --"Gandering."


No comments:

Post a Comment

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