15 Models and a Trickster
Updated: Oct 8
Not unusually, it started with a question that was more of an itch.
I’d been playing with generative image and video models for a while, and one thing kept nagging me: when people talk about “AI” making an image, it’s easy to imagine a singular thing called AI sitting behind the curtain, waiting to produce an output. But that’s not how it works.
There are models. And lots of them. They have different training histories, architectures, and tendencies. Different ways of interpreting the same words. So I wondered what would happen if I gave a bunch of them the same instruction.
One prompt. Fifteen models.
The prompt was, obviously: “half, unusual.”
That phrase is a recurring disturbance in my work. It started a bit accidentally, but I’ve used it often enough now that it has acquired something of a mythology of its own. Unwittingly building lore ..,
Anyway, this time I decided (again) not to tell the models what I was looking for. I just asked them to make something.
Fifteen answers
The interesting thing wasn’t simply that the images were different. Of course they were different; that was kind of the point.
What interested me was how they were different. Each model seemed to have its own visual instincts, its own little collection of assumptions about what an image ought to contain when confronted with the same deliberately vague instruction.
Some results were strange. Some beautiful. Some … well, funny. Some were surprisingly literal. Some seemed to take the prompt somewhere I hadn’t considered. And some models have begun to acquire distinct personalities in my head. Flux 2, for example, is increasingly becoming what I think of as the Trickster’s photographer: clean, clear, uncluttered images that somehow still manage to be wonderfully odd.
And then, looking across the collection, something else became apparent. Patterns began to appear. They weren’t necessarily patterns that occurred in every image, or even in most of them. They were more like recurring gestures: things appearing, disappearing, mutating, becoming other.
There were faces, bodies, objects that didn’t really behave like objects. There were moments when one thing seemed to be turning into another. Basically, there was a gnawing feeling that the images weren’t particularly interested in behaving themselves.
I hadn’t asked for any of that stuff. Which led to me thinking about the Trickster.
The Trickster
The Trickster is a slippery archetype. Across traditions this tricky figure turns up as fool, joker, thief, shapeshifter, boundary-crosser, disruptor, mischief-maker. The Trickster breaks rules, but isn’t necessarily trying to destroy the system. Because sometimes the disruption is precisely what makes something new possible.
That seemed strangely appropriate. This exploration really struck a chord with the halF unusuaL project as a whole.
I definitely wasn’t trying to make a Trickster. Well, I don’t think I was, in any case. I was trying to see what happened when different generative systems were given the same ambiguous instruction. What makes them tick …
But the more I looked at the results, the more the experiment itself began to feel Trickster-like. It just wouldn’t settle down. Every time I thought I had spotted a pattern, another image or apparent direction would undermine it. Every answer seemed to generate yet another question. Oh, and then there was the mouse.
Keep an eye on the mouse
There is a tiny mouse somewhere in the video. I’m not going to tell you exactly where. Partly because finding it is more fun than being told about it, but also because the mouse became a useful illustration of what had happened to the experiment.
I reached a point where I stopped simply looking at the images - I started looking for things. Once that happened, the experiment changed again. A detail that might otherwise have been insignificant became significant because I was watching for it. The viewer does the same thing. Once you suggest that there is something to notice, attention starts manufacturing relationships.
Is that another little Trickster move? The thing doesn’t seem to need to be important. You just have to notice it. 🐁
One prompt isn’t one meaning
Given all the above, there’s another reason I like the experiment. The phrase “half, unusual” is almost comically under-specified.
Half what?
Unusual how?
There isn’t really enough information there to determine a single answer. And that’s precisely what makes the experiment possible. Each model has to fill in the gaps, which means that the outputs aren’t simply illustrations of the prompt; they are interpretations of it.
Fifteen models get one instruction. What follows is fifteen different attempts to decide what that instruction means. And that might tell us something useful about the phrase “AI creativity”.
The model isn’t simply translating an idea into an image. It is participating in the process of deciding what the idea could have been. And that crucial distinction really interests me. Both as halF unusuaL, but, honestly, as me too!
An experiment that became something else
So, I started the experiment because I wanted to compare models. But I ended up making a video about a Trickster; which has become a fairly typical trajectory for my ‘crazy’ experiments.
The original question gets answered, but in the process of sorting that one out another stray question saunters in off the street. This is so one of the things I enjoy about working with generative systems. You can set up quite a precise procedure and then still end up somewhere you didn’t anticipate.
I’m tempted to describe that as randomness, but I’m not sure that’s quite right.
The process isn’t random in any straightforward sense. The models have mind-bending amounts of structure behind them. The prompt serves to constrain what they do. The very nature of the model constrains what it can do. The settings then constrain it further. And then I come along and select, arrange, edit and sequence the results.
The final video is not something simply born, complete, out of a machine. It feels more like the residue of a conversation between several different kinds of machinery, language, images, decisions and accidents; both human and non-human.
And somewhere in that conversation, the Trickster has appeared. Or have I just noticed it? I’m not entirely sure which.
15 models and a Trickster
So there you go. That’s the experiment, and this is the recipe:
15 x models
1 x image prompt
1 x collection of wildly different interpretations
1 x Trickster moving between them all!
I don’t know whether the Trickster was hiding in the models, in their training, hiding in the prompt, or hiding in the selection process. Were they hiding in me?
It doesn’t matter. The useful thing about a Trickster is that it generally doesn’t seem to leave you with a neat answer. It leaves something slightly displaced.
Something you weren’t expecting. Something unusuaL? It’s something that makes you look twice. Look again. 1
So, if you watch the video, keep an eye on the mouse. 🐁
I wouldn’t assume you’ve seen everything the first time.
______
1. See footnote 85 😁
Comments