8 AI generated images of buildings, landscapes with people, all with the same black dot in the center.

Dots and eyes

Feb 16, 2024 | AI Collaborations, Art, Photoshop

From humble beginnings

It might look like I haven’t done much recently, I realise that the obvious christmas post is obviously a couple of months old now. And I’m not going to do anything much to convince that I’m not just spending my days playing on the xbox and/or gazing out to sea. You’ll have to just trust that I’ve been busy. Because today I’m going to go back a little bit to something I started working on really early on in my AI adventures.

I was curious what would happen if you gave Photoshop’s Firefly very little to start with and asked it to generatively fill a large area. I started with a simple black circle in the middle of a large square. This is what I got.

Quite a lot of variety in the generated content. A lot of figures in landscapes, a lot of architecture, some nature and some abstracts. Varying degrees of convincing. It’s interesting how the wide variety of subject matter means you might not notice the central dark circle tying all the images together, but I wonder if that one spot of consistency would be picked up on a subconscious level. Later I played with more dots, in patterns and with prompts, I will come back to those in a future post. 

For now, the next experiment involved an AI generated eye, expanded with no prompts. Here’s the results in super fast form. 

These, I felt, were a lot more interesting. The extra detail in the visual prompt meant the AI had to work harder to figure out what to do to expand the image: it had to figure out what it was looking at before it could decide what else to create. A lot of the results show recognition that the prompt was an eye, but loads of images are completely unexpected. Again there is quite a bit of architecture, fountains and lighthouses making some sense, but also some really fabulous abstract, painterly and sculptural creations. The problem with these is that the eye is so obvious to us organic-intelligences the join jars. So I went back to each of the generated images and regenerated a patch over the eye. Taking the surrounding visual information Firefly covered over the original eye with whatever it thought made the most sense. Which is unfortunate, because what I really wanted to do was still have an eye in the eye place, but have it match the surroundings better, to blend in. Still, the final set of images is kind of fabulous. Much like when I was making the Twinsies images, I came away from this experiment wishing I had the space (and skills) to recreate these ideas in paint.