Stochastic Jué (觉)
Attention, Divergence, and the Structure of Generation
2026
Medium
Immersive installation
Year
2026
Diffusion model, eye tracking, EEG, Gemini API, Touchdesigner
Exhibitions
MAT End of Year Show, 2026
Collaborators
Shaw Xiao, Iason Paterakis, Spike Rao
Stochastic Jué (觉) is an immersive experience that reveals the hidden dynamics of generative models through attention and perception. By combining machine attention with human gaze and physiological signals, the system introduces divergence as a form of intervention, allowing viewers to probe and reshape the generative process. The work frames perception not as observation, but as an active window through which stochastic structures become visible.
觉 [the Chinese character for perception, awareness, awakening] frames this exchange as irreducibly bilateral. Stochastic Jué asks what happens at the boundary between these two black boxes when human attention becomes a probe, and the structure of generation becomes, briefly, visible.
Stochastic Jué (觉) places the human perceptual apparatus in direct dialogue with the internal dynamics of a generative diffusion model. Where diffusion is typically experienced as a black-box prompt-in, image-out, this work opens the process from within, exposing the cross-attention fields that structure how the model spatially grounds meaning across generation steps.
The installation introduces eye-tracking and EEG as perturbation signals. As a participant observes a generated image, their gaze coordinates and neural attention state are recorded and compared against the model's own cross-attention maps. The divergence between human and machine focus becomes the operative signal, which is fed back into the diffusion process to destabilize, redirect, or reorganize generative structure.
Stochastic Jué // Gaze-weighted diffusion layers dissolving into a particle field
Stochastic Jué // Memory fragments rendered as layered planes
Particle stream entering the diffusion volume
Layered frames // eye-tracking gaze point in blue
Diffusion planes in depth, viewed from the side
TouchDesigner network // attention map pipeline





