I made an entry to Claire Silver’s AI art contest.

Check it out and maybe show it some love over at Twitter so she notices it.

x.com/simonjaquemet/status/1798116504845615499?s=46&t=gwEzwynRGvj-R3-iMgYyLQ

Here the lengthy description:

Here's my submission: QUALM KINGDOM #ClaireAIContest

The actual piece is a web app. It has only one button: «Generate»

qualm-kingdom.rtgen.ai

Attached some images inside the low limit that Twitter allows. The idea is that the project consists of infinite images.

OK... What is this about:

My father became obsessed with photography after he retired. He took thousands of images of rather uneventful places around Zurich in Switzerland where he and me live. His images had a very understated humor in how he captured the mundane. He shared it with me and we discussed. Inspired by the contest I asked him if I could train a model on his images. He agreed and I trained an SDXL checkpoint on around 50 photos and the resulting model worked well in capturing the essence of a Swiss (probably similar to Germany or Austria) city and surroundings.

Sometimes I get an eerie feeling looking at the somewhat boring tidiness, perfect organization and wealth that surrounds me. When I see images of Ukraine, where I was several times and have friends and which is a short flight or a manageable car-drive away. Or worse the middle-east, which leaves me totally helpless in what to even think or feel I get something like survivor's guilt. The wealth and stability suddenly seems fragile and maybe it is us causing all this, or at least not doing enough or anything.

I took the model trained on my father's images and added 'war' and 'destruction' to the prompt. The result is a stream of ephemeral pictures of an imaginary war-torn alpine western European country.

One of my main interests in AI art is to create a concept and code for endless and unique, fleeting images.

Tech: This runs on a React frontend doing the whole prompt logic and calling a ComfyUI backend on Runpod Serverless. The Comfy Workflow is rather simple with a low-res image generated by my trained model and then being upscaled by RealVisXL-turbo using the great PAG (Perturbed Attention Guidance). Be patient. It still takes around 1 minute per image.

I learnt a lot. I wanted to do much more: like a page that would show an endless slideshow of new images and sound but in the end I met my limits (and ChatGPT was down today when I needed it)... Maybe will update along the way