early stop

A diffusion model builds an image by removing noise in steps — typically a few dozen — with the picture arriving out of a grey static that gets less random each pass. That static is Gaussian white noise, and neither half of the name is incidental. White: uncorrelated, no relation between one scale and the next. Gaussian: its values fall in a bell curve, which at a given power is the most levelled distribution there is. The model was trained by adding Gaussian noise to photographs in known amounts and learning to predict what had been added, and that choice is what makes the method tractable at all, since Gaussians stay Gaussian when summed and the training can jump straight to any point in the corruption schedule. See octaves for why the distribution and the spectrum are separate questions. The operation at issue is halting the process at roughly a fifth of its steps and keeping what is on screen.

The claim to be careful about: this does not interrupt a process that would otherwise have arrived at a finished state, because on individuation's account no terminated state was coming. What the halt exposes is a real incompleteness that a fully denoised output would simply have concealed better — the same operation that reveals infrastructure by breaking it, and the rupture that semiosphere says switches meaning on.

What becomes visible follows from what denoising actually is. Correlation gets installed coarse-to-fine, so an early halt yields a field organised at large and middle scales and still uncorrelated at small ones — see octaves. Hence the specific look: interference patterns and banding where middle-scale structure has arrived without fine structure to break it up, surfaces reading as waxy or corrugated for the same reason, and regions where a category is present as a tendency without having settled into any instance of it. In the terms of the gap this manufactures a joint rather than a hole, and what it yields is best described as a phantasma: a figuration that gives itself as non-present rather than an image with parts missing.