Art by AI? Some Initial Experimentation
#post
This post originally appeared on Blog 2.0
It seems like that there has been a lot of hype about OpenAI’s DALL·E 2 that’s been generating incredible art and I have always wanted to give it a try. Though understanding the inner working of this generative model might need some reading, it doesn’t hurt to play around with it before reading all the relevant papers.
Unfortunately, I am still on the DALL·E 2 waitlist. The closest to DALL·E 2 I can get seems to be DALL·E mini available on Hugging Face. Let’s throw in some prompt and see what we get.
Hmmm… the model clearly understands some of what we want from it. But the quality is probably not good that good, at least when the outset of this exploration was to see if we can make art.
Disco Diffusion if probably a better way to go if I juste wanted to see some art being made. After all it was designed for it. It uses a diffusion model just like DALL·E but is made much more accessible — takes little effort to configure and run it on Google Colab. To run the current version (v5.6), just go to this Colab notebook (assuming you have Colab access). Instructions can be found within the notebook. I wonder who’s behind Disco Diffusion—could not find a paper talking about how it works…
Anyway, I left most setting to default and run 5 batches. It looked much more like art. Here is what we got (with manual post processing). You can guess what the prompt was.
Okay, they are quite interesting to look at. The other ones weren’t as great out of the box. We’ll take it over and see what we can create with the material. Quite tricky getting the shape right and the colour matched!
I guess I shall go read the DALL·E paper. Peace out.