I recently used an image-to-text generator (CLIP Interrogator 2.4) paired with a text-to-image generator (Stable Diffusion 2.1) to recreate some of the great historical art. You can see the results of that experiment at AI Art Forgeries. Then I decided to see what this process would come up with if I asked the image-to-text generator for a negative prompt, that is, a prompt that would produce something as different as possible from the original art. You can see the results of that experiment at AI Art Inversions I. Finally, I decided to use the original prompts from the image-to-text generator as negative prompts for the text-to-image generator. I will present the results of that experiment here. I used the same 68 pieces of art. I then submitted these as negative prompts to Stable Diffusion 2.1 and accepted the first output.
Here are the results (in the same order as before).
Prompts were generated using CLIP Interrogator 2.4.
Illustrations were drawn using Stable Diffusion 2.1.