Undertrained GANs — Matt test print (1080)
A 1080p study from an undertrained generative model: latent drift and half-formed structure treated as print material rather than failure — the model's uncertainty made visible at scale.
This piece comes from the same instinct as the early COVID StyleGAN work: staying with models that have not fully converged, and treating their artifacts as aesthetic material instead of mistakes to train away. Here the output is framed as a test print — something you would run at full resolution to see whether the noise, moiré, and collapsing geometry read as texture, rhythm, or narrative when the pixels are large enough to breathe.
Undertraining is not only a technical state. It is a compositional one. The network is still guessing; the latent walk reads as hesitation made spatial. That hesitation is often more interesting than a polished final frame, because it keeps the viewer aware that the image is provisional — that another step in the walk would have dissolved it differently.
Latent portrait walks from the same summer 2020 series (including the Henry walk) live on the Early Covid GANs page alongside the moiré study.
Matt test print — 1080p capture (H.264)