An experiment is a controlled procedure carried out to test a hypothesis, 
discover an unknown effect, or validate a scientific theory.

Corollary: wave function collapse, or how does a Drake’s nature change depending on how we observe it?

OK. Obviously, that is a bit of a joke. But it is also a pretty good introduction to what happened here.

When I ask GPT Astra to generate an image, I roughly observe the following process:

1. It interprets and reformulates my request: “I will create X, keeping the characteristics of Y, while doing Z.”
2. It then sends that intention to ImageGen: a second level of interpretation, a second black box.
3. ImageGen returns an image.

And as in a scientific experiment, a question soon starts to emerge: what can be reproduced?

Taking notes on my method, I realised I was handling the same components as an experiment: variables I could change, others I could observe, and a whole balance to find between the two.

Independent variable: The factor that the researcher deliberately changes or manipulates.

In my case: the reference image, its role, and above all the semantic framing I give it.

Dependent variable: The factor that is measured or observed to see how it responds to the changes.

In my case: how the model reacts when I change one variable rather than another, what drifts, and above all what persists from one generation to the next. In a regular conversation, you iterate, but the logic eventually gets lost in the history, and then disaster strikes: it starts to “drift and slop”. You have to remember what you changed, which reference was doing what, and so on.

This is not a new problem. I was already having fun explaining composition variables to GPT-4o with my students, and my first notes on variables in diffusion models date back to 2023. In AI years, it feels like the last century.
See the Playmobil workshop

I keep revisiting my experiments as the models improve and my own understanding of this practice develops. You could call it co-creation or AI-assisted design. We have been talking about DTP (desktop publishing) since the 1980s, after all. A little context and perspective never hurt.

For djiiz.art, I needed a thread to connect my explorations.

I had this illustration I made in 2021: an ape sitting in front of an old Macintosh. I still liked the image. The Macintosh, an early piece of hardware that became almost iconic in the imagination of Silicon Valley designers, still made sense.

The ape, a little less.

Introducing myself as a big primate while presenting my latest projects was perhaps not exactly the best personal branding possible.

So I tried something very simple: mixing the images.

ref1:image of the ape in front of a **macintosh**
ref2:image of the **red drake**

gen image : a 3d render clay illustration of the **red drake** sitting in front 
of the **macintosh** - same pose - no bg alpha transparency

negative : no fire, no rim light

The Drake kept enough of its identity to become something else, while taking on the composition and intention of the original image.
I had my mascot.
Except, of course, I did not stop there.

Once the principle worked, I started changing the variables.

A pose. A profession. A material. An expression. A scale. Sometimes even the entire nature of the object.

The Drake became a filmmaker, skater, data analyst, alchemist, construction worker, plush toy, figurine, pixel art, spider and probably a few things I have already forgotten.

About forty variations later, the experiment had gone well beyond the initial test.

These characters are now scattered throughout djiiz.art. They accompany the projects, illustrate ideas and contribute directly to the site’s art direction.

What had started as an image generation experiment had become a visual system.

Win once, you are lucky. Win every time? There is a trick!

Producing one good image is one thing.

Producing forty without gradually losing the character’s identity is another.

The more complex my experiments became, the more the regular conversation showed its limits: references piling up, instructions changing, abandoned branches, variations that were hard to find again…

And above all, that eternal question:

what did I change between this image and that one?

That is exactly where nodes became interesting.