Our TouchDesigner Pipeline

TouchDesigner is the spine of most of our real-time work, but the pipeline matters more than the tool. What follows is how we actually build, taken from shipped projects rather than from a tutorial.

It starts in the shader, not in the network. For the 100 Violins concert the entire look lives in one GLSL TOP. Flat colours, slow gradients, no figurative imagery, because the piece sits around an orchestra and must not compete with it. Writing the look as a shader rather than as a node chain means the whole system has one source of truth, and a change to the palette is a change in one place instead of sixteen.

TouchDesigner then does what it is genuinely good at: routing, mapping and live control. Eight pillars, two faces each, sixteen surfaces, with the inner faces on a warm range and the outer faces on a cold one. That split is a parameter in the network, not a second export. NDI carries the output to the media machines.

Performance discipline is not optional at this scale. Cook on demand wherever possible, and be careful with viewer flags, because an open viewer forces its operator to cook whether or not anything needs it. Simulation resolution stays small and gets upscaled late. A composition that runs at sixty frames on a workstation and stutters at the venue has usually lost its budget to nodes nobody is looking at.

For interactive work the pipeline extends past the computer. Djerba Explore ran on Arduino UNO boards with HC-SR04 ultrasonic sensors, an Enttec USB DMX Pro Mk2 as the DMX interface, PARELED fixtures for light, and a 400 by 600 pixel LED panel for wall pixel mapping. Touching the olives inside a jar closed a sensor loop that triggered both the wall and a sound layer. None of that is exotic hardware. It is ordinary components wired with intent.

Audio input deserves its own note. At Djerba, nine microphones spread at different heights of the room fed a single audio interface, and that signal drove a real-time datavisualisation projected on the floor. The room listened to itself. Treating microphones as data sources rather than as recording gear opens a category of work that no video file can produce.

We build a browser-based generator alongside the TouchDesigner project on most shows. It serves two purposes. It is a preview that can be shared with a client who does not have TouchDesigner installed, and it is an NDI fallback if the main machine fails. Building the look twice sounds wasteful. It costs an afternoon and it has saved shows.

The last part of the pipeline is written, not built. A setup document describing the network, a runbook listing the order of operations, and an inventory down to the component. On a project with sensors, DMX and multiple machines, the network diagram is as much a deliverable as the visuals.

The tool is never the point. The point is a generative loop that runs on its own, that a tired person can restart, and that produces something specific to the room it is in.