posecode-render

posecode-render

Renders a posecode-parser IR as an animated 3D human figure with Three.js: forward kinematics plus ground-lock CCD IK, live in the browser at 60fps. Pass characterUrl to show a realistic skinned character (any Mixamo-rigged GLB); without it — or while it loads, or if it fails — a procedural athletic figure renders instead, so the scene is never blank. Either way, a capsule self-collision pass keeps limbs from passing through the body.

Part of Posecode: a kinematic-motion protocol LLMs can write, rendered as text-to-motion 3D animation.

Install

npm install posecode-render posecode-parser three

Usage

import { parse } from "posecode-parser";
import { createViewer } from "posecode-render";

const canvas = document.querySelector("canvas")!;
const viewer = createViewer(canvas, {
  autoRotate: false,
  // Metric grid, load origin, live +Z facing arrow, and authored travel path.
  // Enabled by default; disable it for a clean presentation-only embed.
  floorGuide: true,
  // Optional: realistic skinned character (Mixamo bone naming). Omit for the
  // zero-asset procedural figure.
  characterUrl: "https://posecode.org/models/xbot.glb",
});

const { ir } = parse(myPosecodeSource);
if (ir) {
  viewer.load(ir);
  viewer.setLoop(true);
  viewer.play();
}

viewer.onPhase(({ phaseName, cue }) => {
  console.log(phaseName, cue);
});

const floor = viewer.getFloorGuideInfo();
console.log(floor?.gridStepMetres, floor?.hasTravel, floor?.waypoints);

const solverWarnings = viewer
  .getConstraintDiagnostics()
  .filter((diagnostic) => !diagnostic.pass);
console.log(solverWarnings);

cue is optional display-only coaching text. It never changes the pose, timing, range checks, contacts, or collision solving.

Constraint diagnostics name visible heel/toe contact errors, sole tilt, grounding-versus-ankle-ROM conflicts, and residual overlaps from the renderer's bounded self-collision pairs. They measure the post-solver procedural driver, before optional skinned-character or mocap surface reconciliation, and report outcomes without changing the authored motion.

No GPU, no diffusion model: generation is a fraction of a cent of text, and rendering is plain forward kinematics.

License

AGPL-3.0-only. A separate commercial license is available for closed-source product use; see commercial licensing.