Collaborative Movement Study for Realtime AI and Human Improvisation
What is it like to co-create with AI in an iterative feedback loop, where the AI generates thematic visuals in response to the movements of humans who, in turn, improvise dance in response to the visuals generated by the AI?
2025 Mesa Arts Center, 22ft x 12ft
Built in TouchDesigner, this real-time Stable Diffusion experience treats the body as both input and interpreter. Instead of typing instructions into a prompt box, the human prompts with movement. The system isolates the figure from its background and lets gestures, posture, and proximity become the prompt itself. Each motion folds into the generative stream: a hand sweep remaps color fields, a turn of the torso dissolves one leaf shape into another.
Instead of “AI art” as static image, a relational aesthetics emerges from the visuo-somatic negotiation—algorithmic inference converses with human movement. It replaces linguistic precision with the preverbal ambiguity of embodiment. The interface is no longer textual but proprioceptive; what is generated arises through relation instead of command. In this sense, the work reframes “prompt engineering” as an act of listening through motion, a choreography of mutual influence where body and model are collaborators composing in real time.
Running a hyper-compact diffusion model locally on a RTX 4070 gaming laptop, the system sidesteps the usual rendering queues, cloud pipelines and textual bottlenecks—traditional affordances that distance the human artist from the algorithmic expression. Custom scripting compresses inference time so that frames cascade as consequences of motion rather than static, prepackaged endpoints. 24fps image generation—a near impossibility for Stable Diffusion—is treated with some clever post-processing that translates into an uncanny responsiveness, a generative visual articulation that is synchronized with the body and preceding the thought of it.
This immediacy is not just technical bravado; it is philosophical. The real-time exchange collapses the distance that usually governs human-machine interaction and supersedes the wait time of the AI model to “understand.” No more typing the right words to coax an image into being. This dialogue happens at the level of impulse and reaction: AI becomes an environment to be inhabited—a perpetually unfinished state that catches impressions of movement in the act of becoming dance.