MediaPipe adapters

The body-tracking guide shows how to use a tracked body. This page is about the layer underneath: the adapter tools that load the free torinmb/mediapipe-touchdesigner plugin, switch on the pipeline you want, find the engine's output, and hand you a clean, canonical CHOP or matte you can wire into anything.

Reach for an adapter when you want face / hand / body / segmentation data as a stable signal — landmarks in TD coordinates, a person matte, screen-space points for UI — rather than a finished visual.

Get the plugin

The plugin is MIT-licensed and runs on macOS and Windows from an ordinary webcam. tdmcp's package manager can stage it for you:

npx --yes --package=@dpantani/tdmcp tdmcp install mediapipe-touchdesigner

This grabs the MIT release and extracts it under ~/.tdmcp/packages; the adapters resolve the staged path automatically (with legacy and manual fallbacks). See Body & pose tracking for the camera-permission and first-run details.

Timeline must be playing

The plugin's embedded browser only captures while the TD timeline plays. If the data reads zero, press Play.

The adapters

Each adapter loads the engine (idempotent — it reuses an existing one), toggles the pipeline you ask for, and builds a small wrapper COMP whose Script CHOP/TOP reads the engine's JSON output and republishes it as a canonical signal. They probe the engine's output names across plugin versions, so they keep working as the plugin evolves.

Adapter	Layer	What it wires up	Output
setup_body_tracking	1	Pose pipeline → a canonical pose CHOP; optionally builds a skeleton visual.	33-landmark pose CHOP (tx/ty/tz/confidence), hip-centred
setup_face_tracking	2	Face-mesh pipeline → a canonical face CHOP.	468 (or 478 with iris) landmark CHOP, nose-centred
setup_hand_tracking	2	Hand pipeline → a canonical hand CHOP, in world or image space.	max_hands × 21-landmark CHOP, with handedness and screen-space channels
setup_segmentation	2	Selfie-segmentation pipeline → a feathered mask TOP, optionally a pre-keyed RGBA.	mask TOP (+ optional person_rgba for drop-in compositing)
setup_mediapipe_plugin	1	Loads the engine once and toggles Face/Hand/Body/Segmentation together; discovers and exports each pipeline's output path.	the engine's discovered DAT/TOP paths

"Set up hand tracking in world space for up to two hands.""Set up segmentation with a 12-px feather and publish a pre-keyed person matte."

Use setup_mediapipe_plugin when you want several pipelines from one engine in a single call; use the individual setup_* adapters when you want exactly one modality and its canonical wrapper.

From adapter to art

Adapters produce signals; the create_* tools consume them. Point a consumer's source at the adapter's output CHOP/TOP and the rest of the network behaves exactly as it does in synthetic mode:

• Pose → create_pose_tracking, create_pose_skeleton, create_body_reactive (see Body & pose tracking).
• Hands → create_hand_gesture_bus (publishes stable gesture channels) → create_hand_hologram, create_hand_ableton_mapper.
• Segmentation → feed the mask or person_rgba into any compositing chain.

Ready-made recipes

Three browsable templates ship in the recipe gallery — say "list recipes":

• Pose Skeleton (MediaPipe) — the stick-figure skeleton from a live webcam.
• MediaPipe Body Dots — glowing dots tracking every joint.
• MediaPipe Face Overlay — a dot-cloud over live video from the face landmarks.

Each expects the plugin loaded; point its Select CHOP at the matching adapter's output.

See also

• Body & pose tracking — the consumer-side guide and the camera/first-run walkthrough.
• Camera & motion reactive in the prompt cookbook.