[pahn]

I had an art exhibition last week, where I tried to replace myself as an artist with a Henry Ford-style production line (including a moving conveyor belt), run by different AI models to produce my art. The system looks for inspiration in my emails, my bookshelf, surfs reddit and also HN, and then creates pretty fancy installation concepts plus the matching installation views. The 125 artworks it produced within a day and a half can all be checked out here: https://maschinenzeitmaschine.de/p_doppelgaenger_gallery/

Under the hood, it's Python (Flask), mostly using GPT-3 & 4 plus Midjourney (with fallbacks to DALL-E and local Stable Diffusion as using Midjourney programmatically is pretty fragile). I tried fine-tuning models, but as I did not have enough data, I got better results just chaining complex, modular prompts. The conveyor belt was made out of spare parts for treadmills… feel free to ask any questions!

[siraben]

The Midjourney outputs look fantastic! Are those the prompts used in the gallery? Was post-processing involved? Usually there's a very distinctive look to Midjourney images, but these look very believable.

[pahn]

Thank you! I did not use the concepts directly as prompts, but did ask the concept-model to also output a visual description (which was only used internally). This description I then piped through GPT4 (which was much better at this then GPT3), then added some modifiers, and only then sent it to Midjourney. You will often note though, that language comprehension is not really on point – like often it does show the right elements in the image, but not necessarily in the same relationship to another as they are described in the concept...