Turning a Crank to Watch AI Denoise

Diffusion models generate images by starting with pure noise and gradually removing it, step by step, until a coherent picture emerges. In most applications, that intermediate process is hidden behind a progress bar or a single "generate" button. Nobody outside the research lab sees the messy middle states, the moments when an image is half-formed and half-static. Diffusion TV, an installation by Sihwa Park at York University, takes that hidden middle ground and makes it the thing you play with, using a physical object most people remember from childhood: a CRT television.

The Problem with "Just Ask an AI"

Explainable AI (XAI) has a well-documented format problem. Most explanations arrive as 2D screen interfaces, charts, heatmaps, or textual summaries that assume the viewer already knows what a model is doing. For creative applications of generative AI, this creates a gap. The technology shapes how art, music, and video get produced, yet the people using or encountering it often have no intuitive grasp of how it works. Traditional XAI methods treat explanation as information delivery, not experience.

Park situates Diffusion TV within a growing body of work on alternative XAI. Colley and colleagues introduced "tangible explainable AI" in 2022, exploring how physical objects can serve as explanation modality. Ghajargar and collaborators proposed "graspable AI," using material metaphors and aesthetic qualities to support understanding through perception rather than instruction. Hemment and team coined "experiential AI," leveraging artistic practice to mediate between opaque systems and human comprehension. These approaches share a common argument: showing someone how an AI system works can happen through interaction, not just through words on a screen.

Diffusion TV builds on this lineage but targets a specific machine learning mechanism that is unusually well-suited to physical metaphor: denoising. Unlike a classifier's internal weights or a transformer's attention heads, a diffusion model's denoising steps produce visible, sequential outputs. Each step moves from noise toward clarity. The gap this paper fills is making that sequence tangible, continuous, and explorable by a general audience, without requiring any technical background.

A CRT Television as a Denoising Dial

The installation uses a modified 1987 GoldStar CMX-4200 CRT television. The choice of hardware is deliberate, not just nostalgic. CRT TVs had physical controls: channel knobs, volume dials, extendable antennas. These affordances, the physical properties that invite certain actions, are the interface.

When a participant rotates the TV's telescopic antenna, they are controlling which intermediate denoising step gets displayed. The antenna is mechanically linked to a rotary encoder via a 3D-printed connector, producing a continuous signal. That signal maps to an index selecting image and sound pairs from different denoising stages. Turn the antenna one way and the image dissolves into noise. Turn it the other way and structure gradually appears: an animal's form, its texture, its coloring. The system plays corresponding audio at each stage, so the denoising happens in both visual and auditory channels simultaneously.

The tuning knob, the other classic CRT control, switches between three channels: Past, Present, and Future. Each channel contains a curated set of AI-generated animals. The Past channel shows extinct species. The Present shows endangered ones. The Future shows speculative creatures, imagined by ChatGPT o3. Each channel carries a different emotional and ecological weight, and the denoising metaphor gains additional resonance: you are literally turning back time (Past), engaging with a precarious present, or tuning into something that does not yet exist.

Once a denoising sequence reaches its clearest state, the system displays a brief lower-third graphic with contextual information, such as species name, last-seen year, or habitat details. After a pause, the system resets to the noisiest version of a new animal and waits for the next interaction. Content regenerates at configurable intervals, meaning the installation produces different audiovisual material on different days, rewarding repeat visits.

How the Pipeline Works

The system has three main layers: content generation, storage, and real-time playback.

Content generation happens offline in Google Colab. Park used Stable Diffusion XL for images and Stable Audio Open for sounds. For each of the 16 extinct species, 14 endangered species, and 12 speculative creatures, prompts were manually designed and iteratively refined. The Hugging Face Diffusers library ran inference, and custom callback functions extracted intermediate outputs at every denoising step. That means each animal has a full spectrum of images and sounds, from pure noise to final output, saved as discrete files.

Those files live on Google Drive, shared between the Colab environment and a Raspberry Pi 5 that runs the installation. The RPi runs a Processing client program and a Python sensor module. The Processing client handles content playback and interaction mapping. When the antenna rotates, the client reads the encoder value, selects the corresponding image-sound pair from the pre-generated dataset, and displays it. When the tuning knob clicks to a new channel, the client swaps to the corresponding animal set.

The RPi outputs its display signal through a modified HDMI-to-RF modulator. The signal is broadcast on an analog frequency that the CRT TV picks up through its antenna input. This is a key design detail: the TV is not being used as a monitor. It is receiving the content the same way it received broadcast signals in the 1980s. The experience of watching feels authentic to the object.

A magnetic encoder on the VHF tuner detects channel changes. Both encoder values are fed to the Processing client through the Python sensor module. The client also periodically checks for newly generated content on Drive and updates its dataset asynchronously, so the interaction is not interrupted during content refreshes.

Exhibition Observations

Diffusion TV was exhibited at the International Symposium on Electronic/Emerging Art (ISEA) 2025 in Seoul, South Korea, at the Hangaram Design Museum, from May 23 to 29, 2025. Park reports observations from informal audience feedback and video recordings rather than a structured empirical study.

The antenna interaction was generally perceived as intuitive. People figured it out through exploration. Prior familiarity with CRT TVs influenced how quickly people engaged: older visitors who remembered physically adjusting antennas navigated easily. Younger viewers, less familiar with the hardware, sometimes needed more time or expected the TV to respond to touch or voice. Some participants did not fully explore the sequential content or expected additional system behaviors beyond what was implemented.

The tuning metaphor worked as intended for most visitors. People understood the broadcast metaphor, moved between channels, and engaged with the denoising process without explicit instruction. The ecological and temporal themes of the three channels encouraged thematic interpretation, with visitors connecting the content to broader questions about extinction, conservation, and technology's role in imagining futures.

What This Does and Does Not Guarantee

Park is clear-eyed about the limits. Tangible interaction with the denoising process supports intuitive engagement, but it does not guarantee that someone walks away understanding what a diffusion model actually is, how the math works, or why denoising is the chosen generation paradigm. The installation prioritizes experiential understanding over technical comprehension. That is a feature, not a flaw, in the context of the XAIxArts workshop where this was presented. The argument is that for creative and public-facing contexts, emotional and sensory engagement can accomplish what a lecture or a diagram cannot.

The observations were qualitative and informal. There was no controlled comparison with a traditional XAI visualization. Future work, Park notes, will involve structured audience studies to measure how embodied interaction shapes interpretation. The system will also be extended toward real-time or hybrid inference pipelines. Right now, the content is pre-generated, which means the installation cannot respond to novel prompts in real time. Moving toward live inference would enable more flexible generative behavior but introduces latency and computational demands that the current setup avoids.

Practical Takeaways for Developers and Artists

For developers working with diffusion models, Diffusion TV offers a concrete design pattern: expose intermediate denoising steps as explorable states, not as a hidden progress metric. If you are building a creative tool or an interactive installation around generative AI, consider making the process the interface, not just the output.

The hardware stack is reproducible. A Raspberry Pi 5, a Processing client, encoder sensors, an HDMI-to-RF modulator, and a CRT TV. The content pipeline uses publicly available models: Stable Diffusion XL, Stable Audio Open, the Diffusers library, and Google Colab for inference. The animal data was sourced from IUCN Red List and World Wildlife Fund, with speculative species generated via ChatGPT o3.

The multichannel structure, with its temporal and ecological framing, demonstrates how content curation can give technical interaction emotional weight. The denoising mechanic is the same across all channels, but the meaning changes depending on whether you are watching a thylacine fade into focus, a Sumatran rhinoceros, or a creature that does not exist yet. Context turns a technical process into a narrative.

For anyone interested in tangible XAI, the paper points to a productive direction: use physical metaphors that people already know. A TV antenna is not a metaphor for denoising in the abstract. It is a specific, familiar action, adjusting something to make a picture clear, that maps directly onto a specific technical process. The best embodied explanations work because the physical action and the computational process share a structure, not because the installation wraps a screen in wood and calls it tangible.

Read the paper on arXiv