Pareidolia
In progress
Single-channel video — 30-minute latent walk
A continuous traversal through the latent space of a StyleGAN3 model trained on inkblot forms. Nothing is edited or cut; the movement is the model interpolating between points in its own learned space, one unbroken drift.
Pareidolia is the tendency to find faces in wall stains, animals in clouds, meaning in noise. It is usually described as an error — the recognition system firing on insufficient evidence — but it may be closer to the system working as built. Recognition has to run ahead of certainty to be useful.
A Rorschach plate is that error weaponized: a stimulus engineered to mean nothing, so that whatever appears came from the viewer. This one has been passed through two machines. The training data is a custom dataset built using Midjourney — synthetic plates, prompted and curated toward a specific quality of ambiguity, already one remove from the thing itself. That dataset then trained a GAN, which learned its statistics and generates from that learned distribution.
The training method matters here more than it usually does. A GAN is two networks in opposition: one generating forms, the other judging whether they read as real. Every frame was shaped by a machine repeatedly asking does this look like the thing — the same question the viewer asks, involuntarily, within seconds. The recognition apparatus that made the images and the one watching them are running the same operation on the same ambiguity.
Neither finds anything that was put there. The plate is still empty. The duration is the point: long enough to start reading things into it, and long enough to notice yourself doing it.
Colophon
Custom dataset built using Midjourney, upscaled, and trained locally on an RTX 4090 with a StyleGAN3 fork patched for non-square synthesis (1152×2048 portrait), NVIDIA’s original implementation being square-only.
Hot Bit Soup, 2026.