Dream Pipeline

The screen continuously outputs imagery generated from live speech. Spoken input is captured by a microphone and streamed through whisper.cpp for transcription. A Python script segments the text over time and uses a hand-built dream dictionary to assign thematic and emotional tags, expanding them into prompts sent to TouchDesigner to drive a real-time diffusion system. The work does not aim for semantic fidelity; instead, speech rhythm, pauses, and volume steer how the visual field evolves.

A camera simultaneously tracks movement in the space, leaving brief afterimages and traces in the generated output. 

These traces alter the diffusion behavior by shaping where and how noise is injected, allowing bodily motion to enter as a perturbation rather than a depicted subject. Fogging and glitch layers are added at the output stage, with glitch intensity mapped to volume—louder voices produce stronger ruptures, quieter speech yields subtler distortion. The result is a continuously shifting perceptual field instead of a stable representation.