Thursday, 3 September 2020

Artificial Ignorance

Last month I was using Artificial Intelligence (AI) to clone my style and generate new fake bad Krrrl drawings by computer.  Then I wondered how to subvert this technology -- in Artificial Ignorance -- to make entirely new and unanticipated images.


Fake Drawing!
I did not draw this bird!
(though it was generated partly in my style)

  
AI further
 colored and altered this fake drawing  --
approaching Sarina Brewer



Half Bird/Half Krrrl drawings --
 the "latent spacewalk" video 
from Runway ML


The same video above,
after it was run through a different Runway ML program
MUNIT: Flowers to Faces


Another interesting bird morphing in Runway ML
done by Propagation.house,
not me




PROCEDURE

I used Artificial Ignorance to create images that were half bird, half Krrrl drawings.

Initially I had "trained" my first "model" on birds -- in 3000 steps -- using StyleGAN2 in Runway ML, after uploading 209 high resolution images from my book.


I first trained my drawings
on a model of birds,
in StyleGAN2 on Runway ML


Holly Grimm later suggested that I make a "model" blending my drawings with the birds.  I figured out how to go back way earlier in the training process -- to the 1000 steps checkpoint -- where the images were half bird, half Krrrl drawings.  After reverting to that early point, I downloaded the hybrid images, and the mutant morphing video above, and posted those chimeras to this blog entry.

The YouTube tutorial below gave me an insight on how to do this.



After clicking the "Change" link,
the next screen allowed me to go back
every 500 steps in training,
so I could get the half bird/half Krrrl drawing model


This arrested training method is like horizontal DNA transfer , but between two art styles instead of between plants.  Does that not make this approach in fashion with contemporary thinking?


FAKE HYBRID DRAWINGS
Half Bird/Half Krrrl Drawings

I cherry-picked some of the best hybrid images, half bird/half Krrrl drawings, and posted them below.  I also pushed them further in another AI program, Deep Dream Generator.


Half bird, half bad Krrrl drawing,
100% fake




Deep Dream Generator










A collection of images like this
might make a good indecipherable book
like the Codius Seraphinianus






Put a Bird on It!








Less bird, more figure



Less bird, even more figure







Also altered in Photoshop Elements



The two animations below were made online in a Google Colab program that uses AI (artificial intelligence) to impart artificial parallax to still images, and then wiggles them around.




More Fun with PNGs
From a digital collage I made last April


STYLE B(L)ENDING

I wondered if we can blend two AI "models" to spawn something completely different.  

I want to abuse StyleGAN2 and other AI programs to generate things that we have not seen before, or perhaps not even imagined.  Something like AI Surrealism.
***

Derrick Schultz has a YouTube tutorial explaining how to do this:

YouTube tutorial:  Demo: How to mix models in StyleGAN2

"So basically what you are going to do is, the process you are going to train one model all the way through, and then you are going to start training another model, and we are just going to look at the very early iterations where I get this mix, where it's kind of a mix of both things..."

I think this is essentially what I did in Runway ML, when I trained my 209 drawings on a model of birds.  I arrested that training process early, at the halfway point between bird and Krrrl drawings.

**

However I could have trained on a model that I created in Runway ML, rather than accept the pre-loaded bird model that they gave me. I could have trained on a "model" of marsupials, buildings, science fiction monsters or WikiArt, in order to surprise myself with even stranger hybrid results.

In fact, I did just that, and in the most incestuous manner -- by training on the custom models of my bad drawings.  After creating the initial model from my drawings (209 drawings from my book) in Runway ML, I trained a second model (1456 more Krrrl drawings) on the first model.  So it looks like this kind of style blending can be done inside of Runway ML, without resorting to coding, as Derrick Schultz does in the tutorial.

I could also go in the completely opposite direction.  Instead of arresting the training process, I could just train infinitely inside Runway ML.  Later I uploaded 348 long drawings (usually from 3 hour poses), and trained that data set on the second model to create a third model of bad Krrrl drawings.  I basically trained the third model on two previous iterations of my drawings -- and apparently could train a fourth and a fifth model the same way, like mirrors reflecting mirrors, ad infinitum, to create the ultimate inbreed model of my drawing style.


***

PKL DOWNLOADS

Below are links to PKL files of two models based on my drawing style, in case anyone wants to blend or push them further.

  • PKL model based on the longer drawings from my blog (usually from 3 hour poses), generating 1024 x 1024 images -- DOWNLOAD HERE (364 MB)


Note:  One can upload the PKL file to a Google Colab to generate fake images for free, as well as a "latent spacewalk," as explained in the tutorial below (however there is no direct link to the Colab page)


***

Bend vs Blend

Apparently there is a way to "bend" the style  in AI to get surreal results, as opposed to just "blending" two styles together.  This involves intervening at some layer in the computational GAN process, and subverting the criteria.  I don't understand how to do it, but Derrick Schultz explains how in his YouTube tutorial:



Can I whack the computer while it is training?

"Style/Network Bending" is probably comparable to what King Tubby did in Jamaica, when he "physically hit the spring reverb unit to create a thunderclap sound," and otherwise subverted the conventional production processes, to give that distinct sound to his dub masterpieces...masterpieces... pieces...'eces...es...


***
 
PLAYFORM

The online Playform website seems to do a similar thing.  The video below explains how it blends two sets of images to essentially make a program, a program that will then alter single uploaded images.  It looks like it could create a program to turn any photograph of a human into a fake drawing in my style.



Playform Style Transfer


However it costs money to do this, so I have not tried it yet.  I did start by uploading a bunch of my drawings to Playform:


I started to upload a bunch of my drawings into Playform
to try to create a program to blend styles --
however I did not want to pay money 
until I understood the process a little better


One can sign up for FREE to sketch in Playform, which will transfer the style of your choice onto your crude contour drawing, in essentially real time, at no cost:

Style transfer to your drawing
in Playform


So I uploaded one of my drawings, and drew in Playform to create something unexpected.  Playform is supposed to apply the style of the uploaded image to the sketch I did in Playform:
 
1)  I uploaded this fake drawing to Playform as a style reference:

Input image as style reference


2)  I drew an abstract image in Playform (note the white lines on black background):



3)  The resulting image had color, and didn't look anything like my input drawing:

Interesting AI result in Playform


I think this is the same thing that Scott Eaton is doing  -- as his AI fills in his contour drawings in real time -- though Scott Eaton's figure drawings come out more realistically:

Scott Eaton's AI 
seems to make more successful fake figure drawings


***

Playform has a landscape version of the same drawing/style transfer program:

1)  I uploaded this fake drawing as a style reference:

Style reference from one of my fake drawings

2)  I drew a horizontal landscape-like contour drawing (the middle white line drawing on black):



3)  The result was a good landscape drawing, but did not seem to reference my input style image:

Playform generated a nice landscape
that did not seem to reference
the style of the image that I uploaded

***

AI can now complete your drawings, using GPT?


AI completes cat drawings



***

Unrelated?  Open Processing Art:

Created by Sayama
in Open Processing


CONCLUSION

"Artificial Intelligence" implies a preordained purpose, such as creating realistic mug shots of people who don't exist; or creating realistic portraits of historical figures through ancient sculptures; or creating the next Rembrandt.

However the "Artificial Ignorance" approach aims to confuse things, using that same technlogy in order to come up with a wonderful and unanticipated image, otherwise known as a "happy accident" in the art world.  It is reaching for the image Antegooglewhackblatt, something that even Google does not even know about yet.

I want to use "Artificial Ignorance" to willfully use ignorance to discover an image that I could never have imagined.  This is kind of a backwards approach to creating art, not unlike that of using Brian Eno and Peter Schmidt's Oblique Strategies to overcome creative block.

However ultimately there will be a lot of post-processing involved, in other AI programs and Photoshop, in order to get any decent results that the public would be interested in.

***

 Resources


*****

I liked it when "Artificial Intelligence" (using GPT 2) defined "Artificial Ignorance" in this online computer program that finishes your sentences when you hit the TAB key:

Artificial Ignorance is a form of self-awareness, which is not only a means to self-correct in its own way but also a means to self-correct itself.

Note:  The red text highlighted in blue is the part that was written by Artificial Intelligence.
Posted by at

No comments:

Post a Comment

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

Subscribe to: Post Comments (Atom)