2022generative-ai · computer-graphics · independent

Twin Peaks

A GAN trained on parsed frames from Twin Peaks, distilling the show's recurring palettes and compositional symmetries into entirely new, invented frames.

A GAN-generated Twin Peaks portrait: a distorted figure in a suit, warm orange-brown palette.
01

Overview

A short generative AI experiment usingTwin Peaksas source material, made in May 2022. I assembled a dataset by parsing frames from clips of the series, then trained a GAN on the resulting imagery to explore what the model could learn, and invent, from that visual world.

02

Concept

The GAN revealed unexpected patterns in the show's cinematography — recurring colour palettes, compositional symmetries, and atmospheric qualities — which the model distilled and recombined into entirely new frames that never appeared in the source footage.

A 2x2 grid of GAN-generated Twin Peaks frames, recombining recurring compositions and palettes from the source footage.
A 2x2 grid of GAN-generated Twin Peaks frames, recombining recurring compositions and palettes from the source footage.
03

Reflection

The result is less aboutTwin Peaksspecifically than about what a GAN surfaces when trained on a single, stylistically consistent source: the model doesn't reproduce the show, it reproduces its visual grammar.