Diffusion Local Time, 2024.
This is an exploration of the surrealism of generative AI. A quotidian digital clock is made strange through its face re-imagined in an AI-generated image. View the clock in real time at https://leebutterman.com/diffusion-local-time/
The numerals that signify the time are integrated seamlessly into a visually rich landscape, and the user’s perception is challenged: are they able to discern the time embedded within the randomly-generated and constantly shifting image? Diffusion Local Time pins its core utility on the viewer’s ease of pareidolia, usually associated with seeing faces in clouds.
Diffusion Local Time uses Stability AI’s Stable Diffusion XL 1.0, combined with Monster Labs’ QR code monster as a ControlNet to control luminosity, to render one constant image prompt, with a constantly changing control image of text of the time, rendered in a font for low-vision readers. The random seed for the diffusion changes every hour, to provide visual consistency within the hour and increase variety as the work progresses, and the text eases in and out from minute to minute.
The slowness of the frame rate (several frames a minute) and the subtlety of the text in Diffusion Local Time contrast with the intense speed of the required computation and the legibility of the conventional digital clock face in the modern technological environment.
Diffusion Local Time exercises our cognitive drive to make meaning out of visual patterns, just as its models were tasked with recognizing patterns in the images and texts that they were trained on. Diffusion Local Time lives in a space between 1) Christian Marclay’s The Clock, a video installation supercut of clocks or timepieces over a 24-hour period; 2) procedurally generated artwork and systems music like Brian Eno’s Music For Airports and Terry Riley’s In C, embracing chance in a repetitive structure; 3) the Clock of the Long Now, especially its consideration of longevity and evolvability; 4) nephelomancy, Rorschach blots, LLMs, and other systems of packaging hallucinations into truths.
To minimize the potential for harm, Diffusion Local Time does not include imagery of people or politically significant events, only landscapes embedded with numerical digits. The artwork embraces reusing open-source models and low-power compute, against a prevailing paradigm of renting high-power expensive surveilled machines, or using black-box APIs. The runtime costs for Diffusion Local Time are pennies a day of electricity only, with no internet connection required. Runtime requirements are flexible, and this timepiece can comfortably downgrade to run on a Raspberry Pi.
The source code is publicly available at github.com/lsb/diffusion-local-time and a real-time clock is publicly available at https://leebutterman.com/diffusion-local-time/ .