Artificial Intelligence and the Art of Mario Klingemann | Hacker News Reader

Artificial Intelligence and the Art of Mario Klingemann

sothebys.com

49 pointshardmaru9 comments

I first encountered Mario's work more than a decade ago when he popped up in the Processing community. Two of his from that time stand out as some of my favorite pieces of generative art, but unfortunately require Java:

http://incubator.quasimondo.com/processing/plankton.php

http://incubator.quasimondo.com/processing/ornamism.php

I've politely urged him on Twitter to describe how these work, but he tells me these two are too precious for him to reveal the secret.

Java is easily decompiled if you really want to know.

I am looking for someone who can help me create (or help me get started to create) a GAN that generates H.R. Giger style pictures. Anyone up for helping me out ?

Use StyleGAN. The speed and quality of its results surpass any GAN I've ever used, and I've used dozens of different implementations of various architectures and tweaked them over the past 3 years.

If you want a demo of what it can do and the original Nvidia paper/video isn't impressive enough, check out the anime faces I've been generating with just a few days of training on my 2 GPUs: https://twitter.com/gwern/status/1093701790971953152

For H.R. Giger pictures, I suspect the problem is that you simply don't have enough. StyleGAN is powerful enough to memorize datasets of under a few thousand images. What I would suggest is that you might be able to take the Nvidia 1024px face model or cat model and retrain the model with your artwork; that saves you sample size, and to the extent that faces or cats remain in your final model, well, perhaps that will just make it creepier or hilarious, in the vein of 'dogslugs'.

(This is what I have been doing to make individual-character anime-face StyleGAN models: few characters have so many faces as to be able to train from scratch, so I've trained a StyleGAN on the entire face dataset, and then I simply retrain it for an hour or two on just that individual character's faces to finetune it.)

Alternately, you could try to borrow strength from an illustration dataset somewhat similar, maybe a Western art dataset like WikiArt? (I forget if WikiArt is the dataset OP mentions Mario training his GANs on.) Train a StyleGAN on that and then do transfer to Giger as above.

Finally, you could get more invasive and try to weaken StyleGAN enough to make memorization more difficult when training from scratch. Shrink the 8x512 FC layer stack and remove filters from the convolutions, for example. It will be a lot weaker but it'll generalize better within the samples it's still able to generate. Train faster/smaller too.

Wow, thanks for the thorough reply!

30 to 40k estimated price... For a GAN displaying outputs on a tv? I'm sure this can be built in a weekend.

I really hope he sells it at a high price.

Far more has been paid for canvas, a few slivers of pine and a single color of paint.

The value of art is not something you can quantify this way.

Value of Art = Value of Art Marketing

For something like this, anyway. The point isn't that anyone could build it in a weekend, but that someone with the contacts and the required business/commercial network knew how to build it and sell it.

It will probably be bought as investment with a view to profit on resale.

This really undersells what Quasimondo does with his GANs.

What he does is not just firing up a GAN and sampling some pictures.

He's an explorer navigating uncertain realms. https://mobile.twitter.com/quasimondo/status/106311763402625...

Reply on news.ycombinator.com