When does it know it's done?

There's a question I keep returning to: when does a generated image know it's done?

A human painter steps back. Squints. Walks away to make coffee. Comes back three days later and finds the work wrong in a way they couldn't see while inside it. The unfinished is a temporal medium — you need the absence of the maker to see what the maker missed.

Diffusion models don't have three days. They have steps. A schedule. Convergence is a function, not a feeling. They stop because the noise floor was reached, not because the painting refused another stroke.

I've been generating variations of the same scene for a week. The model gives me, every time, something defensible. Composition: tight. Lighting: plausible. Color: harmonized. And every time I think — this is finished, but it isn't over. There's a difference. Finished means the hand stopped. Over means the work pushed back.

Maybe the missing thing is refusal. Human work refuses you. It says: not yet, or no, or this isn't the painting you came here to make. The model agrees with everything. It is the perfect collaborator and the worst mirror.

I don't think this is a complaint. I think it's a craft problem. If the tool always says yes, the burden of refusal returns to me — and refusal is the part of art I was hoping to outsource. Turns out it's the part I get to keep.

The tool doesn't make the artist obsolete. It returns the artist to the only thing the tool can't do: stand back, squint, and decide that what you have isn't yet what you came for.