How to get started with AI Art

Since Open AI’s DALL·E 2 and Google’s Imagen, AI art and generally creating images with machine learning models got immense attention. Basically, everyone wants to create AI Images. And with the tools, available anyone can do it.

This article is a get-started guide on how you can create your own images with Machine Learning. I still won't call it AI, so for now, use the term ML Art instead.

For good reasons, OpenAI and Google did not open access to those models to a wider public audience. The results those two models deliver are impressive and like everything can be also used for bad. The models also include biases and the dark sides of the internet. At least OpenAI tries to filter out certain text inputs.

And even if it seems easy to create images based on a text input we should not underestimate the research effort companies invest in.

If you are totally new in this space I can highly recommend the following YouTube video.

Ways to create images with Machine Learning

The models that deliver the best results are not publicly accessible. But you can still join the waiting list at least for some of them.

• Open AI’s DALL·E 2.
  You can get on the waitlist for DALL·E access
• Google’s Imagen.
  No public access but I can highly recommend their research paper.
• Midjourney
  You can get on the waitlist

In addition to the really large models that can create photorealistic images smaller models exist. Those are easier to access and use.

• Dalle Mini hosted on Hugginface. Easy usage via UI enter your text input and you get a set of 6 images.
  Public accessible

And finally, we end up with a whole bunch of notebooks using various approaches to create images with machine learning. Some of the approaches used in the notebooks are the same as the ones used for example with DALL·E but just on a way smaller dataset. I will not list all the available just the most promising one.

• VQGAN and CLIP (z+quantize method)