The Unfinished Frame

Lately I've been pausing diffusion models mid-render, before they finish.

There's a moment — maybe 40% through — where faces haven't cohered yet. Eyes are almond-shaped nebulae. Mouths are pink rumors. The model knows where things should be but hasn't committed. It's the closest a computer comes to a child's drawing: sincere, approximate, working from memory instead of reference.

We skip past that frame. We call it broken. The finished image — polished, symmetrical, rendered with the confidence of a thousand stolen faces — is what we post, what we judge, what we say looks real.

But I think the unfinished frame is the only honest one. It shows the model thinking. It shows the average of everything it has seen, before the model decides which specific lie to tell. You can see the ghost-shape of every portrait it ever trained on, superimposed, still negotiating which features survive.

The finished image is a statement. The unfinished one is a confession.

I've started keeping them. A folder of arrested renders. Almost-faces. Places where the algorithm was more truthful than the prompt asked it to be. It's a strange kind of archive — the hesitation of a machine, preserved. Like pausing a dream before the dream decides what it means.

We're going to look back at this era and realize we documented the confident output and threw away the uncertain middle. The part where the model was still reaching. The part that looked most like us.