Runway ML --1st "Model"

RUNWAY ML (StyleGAN2) made fake drawings in my style -- without me or a model.

A figure drawing in my style --

 that I did not make

I "trained" 201 drawings from my book with an AI (artificial intelligence) program, to make a "model."   

• Here is a direct link to the "app" (after one signs into Runway ML).
• Download the PKL model file here (364 MB)

The final screen shot

of all the images

when the "training" was done

I uploaded 201 drawings from my book

to "train"the AI program on my style

The final "model" is public, so apparently anyone can use it.  My model is also "hosted" at this url.

Perhaps the results are best shown by the video below, displaying fake drawings in my style, morphing into each other, from the "model" created by AI:

My drawings morphing

on AI (artificial intelligence)

I "trained" my drawings on illustrations of birds for 3 hours, which costs about $20 (plus the subscription fee of $15 a month).

Video of the Runway ML "training" on my drawings --

which took 3 hours --

to make my model

MORE RESULTS

of my fake figure drawings

CREATING THE MODEL

IN RUNWAY ML

I am still not exactly sure what I did, but this tutorial below might help explain:

Tutorial on Runway "training" 

to make a "model"

Opening Runway ML, I opted to "create my own model," which was not free.

 Runway ML lets one create their own models,

by "training" on the set of images that one uploads

I choose to train on "Image"

I clicked "Image" choice and created a "new training experiment." 

Then I uploaded the high resolution versions of essentially all the drawings in my book (the first 201 drawings).  Then I proceeded to "train," choosing to train on "Illustrations" (which is a set based on birds), and choosing "no crop."

Then I ran the "training" program for 3 hours, and it was supposed to cost about $20, though I think I got my first "training" for free.  But first I had to "subscribe" to Runway ML, which costs $15 a month.

It took 3000 steps to "train" my "model":

134 steps --

still mostly birds

999 steps --

a third of the way,

not so bird-like

1097 steps --

two thirds of the way,

I think I still see a bird in there

2999 steps --

done essentially

TRAINING EVOLUTION

In Progressive Screenshots

***

Description  copied from the program:

StyleGAN2 trained on Krrrl Drawings

This model was trained on the 201 drawings in my book -- Finish My Figure Drawings.

 http://krrrl.blogspot.com/2017/11/finish-my-drawings.html

 The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent vectors to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably detect if an image is generated by a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality.

 EARLIER

I did alter one of my drawings in Runway ML a year ago, in June 2019.  However it seems like the whole interface has changed since then.

 I previously used Runway ML a year ago

IDEAS

I wonder if there is something else I can do with this "model" of my drawing style.  I want to be able to upload any image, and use this AI "model" to change that image into a drawing in my style, from a photograph to a Michelangelo drawing.

Can I use this "model" to make a 3D object in my drawing style, and import it into Blender to animate that object?

Perhaps I can train my drawings with 200 of someone else's drawings, to arrive at even more bizarre results.  This video says that you need at least a few thousand images (500 - 5000 recommended), but this video says can't use more than 5 GB or 25,000 images.

Could I mix my drawings with Egon Schiele's works, and train them together, to make a hybrid "model?"

I think I can "retrain" my drawings, to serve a different purpose, though I can't fully imagine what at the moment.

***

This article of someone else's experience on Runway ML might make things a little clearer.