Motion metrics

Motion graphics fail in two typical ways. Frames hold still too long, and the film reads like slides. Or something jumps with calm on both sides, which reads as a glitch or a cut nobody hid. Motion metrics measure both on small greyscale frames, so a tool (or an AI assistant) can check a piece and say where to fix it.

The measure is the mean absolute change in luma between consecutive frames, on a 0–255 scale, at 192 px wide and 10 frames per second. Thresholds were calibrated on films measured that way.

Measure framesCopy Anchor Link

import { changeSeries, lumaFrame, motionFindings, motionMetrics, MOTION_SAMPLE, MOTION_TARGETS, rasterizeHtml } from "@miraiclip/renderer";

const frames = [];
for (let i = 0; i < durationS * MOTION_SAMPLE.fps; i++) {
  const canvas = await rasterizeHtml({ template, params, widthPx: 1920, heightPx: 1080, timeS: i / MOTION_SAMPLE.fps, durationS, density: MOTION_SAMPLE.width / 1920, scripted: true, software: true });
  const pixels = canvas.getContext("2d")!.getImageData(0, 0, canvas.width, canvas.height).data;
  frames.push(lumaFrame(pixels));
}

const report = motionMetrics(changeSeries(frames), MOTION_SAMPLE.fps, { endHoldSeconds: 1.5 });
// { seconds, density: 0.95, medianChange: 1.6, holds: [{ from, to }], jumps: [t] }

const findings = motionFindings(report, MOTION_TARGETS.motion);
// [{ kind: "hold", from: 4.8, to: 5.7, message: "nothing moves 4.8–5.7 s" }, …]

• lumaFrame(rgba) converts pixels to luma in video range (16–235), so rendered frames match the calibration made on decoded video.
• software: true rasterizes on the CPU. It's cheap at 192 px, and exact: some GPU rasterizers add tile artifacts that would read as motion.
• endHoldSeconds allows an end card to hold, and offsetSeconds reports times within a longer film.

TargetsCopy Anchor Link

For reference, a 13 s motion-graphics launch film measured density 0.99 with a median change of 3.5. Slide-like films with dissolves measured 0.3–0.7 with many holds.

In the React templateCopy Anchor Link

The template's assistant measures every act it composes and the whole film at the end. Each finding comes back with what to change (keep a camera drift running, overlap beats, move further, hide the cut), so the model fixes the act before moving on.