INTERACTIVE INSTALLATION

Diffusion TV puts an AI image model behind a 1987 channel knob

Sihwa Park rebuilt a GoldStar television so that turning its antenna drags a picture out of noise, one denoising step at a time.

Work by Sihwa Park × Luba Elliott × Masaru MizuochiAugust 20263 min read

Sihwa Park's Diffusion TV looks like a thirteen-inch GoldStar television from 1987, and it behaves like one. Static fills the screen. You take hold of the telescopic antenna, turn it, and the noise begins to organize itself: a silhouette, then fur, then an animal, with a sound resolving alongside the picture. Turn the other way and everything falls apart again.

The intermediate frames a diffusion model produces on its way from noise to picture, normally seen only by researchers and usually as a strip of thumbnails in a paper, are here the whole experience, spread across the tuning range of a thirteen-inch cathode-ray tube. There is no diagram anywhere on the piece. There is a telescopic antenna, and turning it is the only way to watch structure appear.

How it was built

Every frame existed before the set was switched on. Park generated the images with Stable Diffusion XL and the sound with Stable Audio Open, running both in Google Colab through the Diffusers library with custom callbacks that captured the output at each denoising step. Those intermediate states were stored on Google Drive, where a Raspberry Pi 5 inside the cabinet collects them and plays them back through a modified HDMI-to-RF modulator.

A rotary encoder, joined to the antenna by a 3D-printed connector, converts rotation into a position in the sequence, while a magnetic encoder on the VHF tuner handles the channels. "I wanted the interaction to feel as though the TV itself was doing something magical, rather than appearing as a visibly hacked device," Park has said.

Diffusion TV — Sihwa ParkDiffusion TV at SIGGRAPH 2026, video by Sihwa Park ↗

From Seoul to Los Angeles

Diffusion TV opened at ISEA 2025 at the Hangaram Design Museum in Seoul, in an exhibition the artist's page records as taking fewer than eleven percent of 355 submissions. It was shortlisted for the CVPR AI Art Gallery in 2025 and selected for the SIGGRAPH 2026 Art Gallery in Los Angeles, where SIGGRAPH's press announcement described it as demystifying diffusion models through a nostalgic CRT television.

Sources

How it was built

From Seoul to Los Angeles

Pixel Agents gives invisible software work a body

StadiView turns a seat number into a place you can inspect

This spreadsheet gives your face an expiry date

We See You.