In Translucence

Nathan Smith
Glasgow, Scotland
grpahicdeisgn.com @nathandavidsmith

Abstract

Viewers walk around a large display positioned as a sculptural object in the centre of the space. Within the sculpture are some exposed computer components, connected to the screen. Visible on the display is the live feed of a camera positioned nearby, pointed at the exhibit and surroundings.

Viewers notice that while they and their surroundings can be seen in the camera feed, the exhibit at which they're looking isn’t visible. The entire artwork is being erased from view in real-time by an offline AI model contained within the sculpture itself. As this camouflage becomes obvious it invites viewers to experiment, forcing the machine to make glitchy mistakes: the machine caught in the act of revealing itself.

Advanced technologies of the past have vanished as they gradually subsumed into everyday utilities: the 'electric light' is simply a light. Though the proposed AI bubble threatens to burst, the tech has nonetheless embedded in the lives of its users. How then might we consider futures for the likes of 'AI-powered journalism' or ‘AI healthcare’? We might describe a technology's propensity to disappear from view as its ‘translucence’. Beyond semantic drift, translucence fundamentally alters the composition of technology, defining how we interact with it. In turn, translucent machines might work to conceal themselves, forming their own meta-translucence.

A referenced paper detailing In Translucence will be published in The Proceedings of the 14th Conference on Computation, Communication, Aesthetics & X (xCoAx) in July 2026. The concept of transparency is introduced in reference to Marshall Mcluhan, while ideas of cultural production and consumption lean on writings from Sean Nixon and others. The creative and critical value of a 'glitch' refers to a definition proposed by Betti Marenko, and Marenko's writing also informs the idea of a meta-translucence.

Technical description

In Translucence is an AI-critical work which runs entirely offline. All processing happens on-device to respect the privacy of viewers who appear in the camera feed. No data is ever sent to remote data centres, nor does the work rely on any external platforms.

A 29" Sony PVM-2950Q CRT monitor is connected to a deconstructed desktop PC through an analogue BNC connection via an HDMI-BNC converter. The PC is surrounded by a custom machined case situated atop the display, and utilises a 12-core AMD Ryzen 9 3900X CPU with 128GB RAM and an Nvidia RTX 4090 with 24GB VRAM. A small offline wifi router links the PC with a small Yi Home 1080p camera which has been flashed with yi-hack-Allwinner-v2 firmware for local network usage.

On setup, the camera is mounted and pointed toward the artwork, and a masked area is defined manually. On boot, TouchDesigner launches four simultaneous instances of ComfyUI, each running a copy of stable-diffusion-1.5-inpainting. A custom Python scheduler treats each stable diffusion instance as a separate renderer, extracting frames from the video feed with timecode and sending to each renderer with a prompt to fill the empty space. The processed frames are then composited back into the masked area of the original footage at the correct timecode (so that the only visible AI processing is within the masked area and the clean feed is untouched) and the composited frames are presented as a live video feed.