Diffusion TV creates a tangible interface for diffusion models by letting users physically tune a CRT antenna to control denoising of AI-generated images/sounds across Past/Present/Future animal channels.
Diffusion TV is an interactive AI art installation by Sihwa Park that transforms a modified 1987 CRT television into a tangible interface for understanding diffusion models. By physically manipulating the TV’s antenna and tuning knob, audiences control the clarity of AI-generated images and sounds, metaphorically enacting the denoising process that underlies generative AI.
The installation features three channels with distinct ecological narratives: Past (extinct species), Present (endangered species), and Future (speculative creatures). This setup foregrounds the generative process over final outputs, offering an embodied mode of explainable AI that allows users to explore intermediate states of AI generation.
Technically, the system uses a Raspberry Pi 5 and custom sensors hidden within the TV to map physical interactions to pre-generated audiovisual sequences representing different denoising stages. This approach avoids the latency of real-time local generation while preserving the nostalgic aesthetic of analog television, connecting the obsolescence of CRTs with the fragility of natural species.
Diffusion TV explores how the internal mechanics of diffusion models can be made tangible and directly manipulable through physical interaction. The system uses a CRT-style antenna as a control surface: users physically tune it to adjust the denoising process that transforms noise into AI-generated animal imagery and sound. By organizing the experience into Past, Present, and Future channels, the work treats diffusion not merely as a black-box generator, but as a perceptible, time-structured signal that can be explored through embodied, hands-on interaction.
Its main contribution is an interactive framework for experiencing diffusion as a controllable process rather than only as a static output pipeline. The antenna metaphor maps a familiar broadcast-tuning gesture onto a latent generative process, allowing users to sense how changing denoising parameters alters visual and auditory results. This reframing is significant because it turns an abstract computational procedure into something spatial, physical, and exploratory, making the generative process more interpretable and engaging.
For a technically literate audience, the work matters at the intersection of generative AI, human-computer interaction, and tangible computing. It suggests a design direction in which users do not simply prompt a model, but directly participate in the unfolding of its inference process. By grounding diffusion in a physical interface with temporal channel semantics, the paper offers a potentially useful model for teaching, prototyping, or creating novel control surfaces for stochastic generative systems.