Ranarh

I've been playing around with AI generated images using Disco Diffusion for a few days and find it very intriguing. DD is open source software that Google Colab provides free access to (as you might expect, you can get nicer features by paying) and can also be run from home. It is a text-to-image generator, meaning you enter a description called prompt, adjust a bunch of parameters to your liking, and watch as an initial colour noise develops into something it "thinks" you described. [Allow me to deviate a bit here: there is a lot of confusion about what computers can and can't do and even more about the capabilities of AI. Right now, they are just dumb machines. Yes, they can produce smalltalk - who can't - and as with any task, if we can use the computer's strength, it shows better results than humans, for example in diagnosing skin cancer, because no human can compare millions of images. But they are just machines. They can't think. They are not alive. We cannot trust them to make decisions for us. Please never forget that, people.]

Diffusion is the math, and Disco is the way it is applied, there are other ways like Latent Diffusion or Progrock (see a pattern here?). If you've used programs like Apophysis or Mandelbulber, you'll feel right at home - change a bunch of numbers, go away for coffee (rendering these things takes time), and come back to find pretty pictures have materialised. Other powerful AI image generators are currently invite only, like the impressive Midjourney and Dall-E, or NVIDIA's beta Canvas that requires quite good hardware.

What I like about the image DD procudes is the density and arbitrary additions taking cues from sources you never knew existed - you literally describe the image to someone else -, it's like being on the client side of illustration for once, and nothing turns out the way you meant to, but is still cool, so you run with it. The same feature of course makes it hard to get it to do what you want (although you can use initial images as a template it will then fill, with mixed success).

As a learning experience for myself as an artist, I think it's very interesting to see how much I need to put to words to describe what I want. I'm not one of the artists who smartly write down a description and keywords for their intended artwork beforehand - maybe I should - so I found I had to really make an effort and think, what kind of image, how dense the detail, the camera setup, art style, maybe influenced by an artist? To DD, it's a big difference if you present the same description for a rainy or sunny day. And everything you leave out it will fill on its own; since the AI was trained with images from the internet, I'm sure the eponymous farm will get a burger drive-thru. That's where many surprises lie, what was clear to you, but you didn't say, may have been left out or filled unexpectedly. Even with a good visual library, I think this experience of recalling it in words is very useful to artists.

Many images cannot (and I guess, should not) be used as-is, certainly not for illustrations, but they are nice starting points and free to use. However, people have indeed managed to make very coherent, detailed images with it, especially landscapes. Output sizes are quite small with a free account; on the other hand, there are tools available that can upscale nicely to workable solutions.

So, don't be afraid of what looks like math, it's really quite harmless, and give it a try to see what you can make it do.

Another tool I have also played with is WOMBO Dream, an app that creates small prompts in premade styles (but also accepts "by <artist>"), with much greater abstraction than Disco Diffusion, and only in portrait format. They too are free to use.

Disco Diffusion: https://colab.research.google.com/github/alembics/disco-diffusion/blob/c8704249481926a126f93741100facbd5471f32b/Disco_Diffusion.ipynb

Most quoted starting guide: https://docs.google.com/document/d/1l8s7uS2dGqjztYSjPpzlmXLjl5PM3IGkRWI3IiCuK7g/mobilebasic

WOMBO Dream: https://app.wombo.art/

Check Discord and reddit for helpful communities; more links in the comments.

See more images than at top on FB https://www.facebook.com/jsl.fans/posts/pfbid0wFKq1NvZC8wduREd2AK3ajmyd6QtJLhHLyHfuziqJCfYd71uZzbyppgTwmr8UG1Gl

or Tumblr if you prefer https://www.tumblr.com/blog/view/ranarh/690131346809569280?source=share

#discodiffusion #digitalart #illustration #conceptart #clip #art #aiart #ai #gan #googlecolab #abstractart #fantasyart #scifiart #tiger #cthulhu #jamesgurney #jeangiraud #birds #trainstation #cyberpunk #fantasyforest #cats #ivanbilibin #aiart