What Stable Diffusion 'knows' about Banksy

I’ve been playing a lot with Stable Diffusion. The most incredible takeaway is that you can generate art in well known styles by simply adding phrases like “a policeman arresting a happy criminal, by Basquiat.” (Yes, I’m exploring what being subversive means within the context of a subversive artist and a new “subversive” medium.)

All the artwork is 100% original, even though it looks just like actual Banksy artwork.

Of course, “original” here means that it was generated by an AI that ingested ALL the works of art by the artist, then created a mathematical representation of what that art looks like to the AI, and then uses that mathematical representation to create new works based on that understanding of the art.

Is that “original?” It begs the question: isn’t all art derivative?

As I played with SD, it was interesting to notice what SD “knows” about Banksy artwork. SD is obviously trained on images that have a caption including Banksy. SD then mimics the artwork. It also adds context to the artwork. For example, the image below mimics the well known “Girl with balloon” artwork, that was originally found on the Waterloo bridge in London.

But, as I generated more and more images, the funniest thing is that SD “thinks” that Banksy art should be in a frame.

Isn’t it ironic that what was originally work to subvert physical public spaces is now understood by the AI to be framed and on someone’s wall.

SD really is confused by how to use cigars and cigarettes. Most of the pictures had them jammed into the eye or random places above the lip.

Corporate lawyers.