Temporal Animation Diff

Temporal Animation Diff

Finds animation discontinuities by rendering a dense deterministic frame sequence, diffing each adjacent frame, and inspecting the highest-delta transitions. This is for motion debugging, not for committed goldens.

Use this before tuning by eye when a clip feels like it snaps, teleports, pops, or has rubbery wobble. The output should name exact frame pairs and phases (p=...) so the code fix can target the real transition.

Loop

stateDiagram-v2
    [*] --> RenderSequence
    RenderSequence --> RankDiffs: adjacent frame pixel deltas
    RankDiffs --> InspectWorst: before/after/diff PNGs
    InspectWorst --> Diagnose: map frame phase to channels, camera, pinning, layout
    Diagnose --> Patch
    Patch --> RerenderSameSequence
    RerenderSameSequence --> RankDiffs
    RankDiffs --> Regression: convert proven bug into a focused test
    Regression --> Cleanup
    Cleanup --> [*]

Workflow

1. Render densely enough to catch snaps. Use at least 120 frames per loop; use 240 when the user reports occasional jumps. Keep viewport, scale, camera, expression, backdrop, and timing identical across runs.

2. Diff adjacent frames. For each pair, compute changed pixels, mean changed delta, score (changedPixels * meanDelta), changed bounding box, and changed centroid. Sort descending by score.

3. Write artifacts for the worst transitions. Save:

   • fNNN.png
   • fNNN+1.png
   • diff_NNN_NNN+1.png

   Put them under build/character_frame_diffs/<label>/ or another ignored build directory. Do not commit them.

4. Inspect before patching. Read the worst before/after/diff images. Decide whether the delta is a true discontinuity or just a large legitimate pose change. True discontinuities usually show whole-body translation, camera jumps, support-foot re-anchors, expression swaps, z-order pops, or limb teleporting.

5. Map phase to code. Convert frame pair to normalized phase: p0 = from / frames, p1 = to / frames. Check channels/keyframes/contact spans/camera curves that cross that phase. For character clips, also compare scene-level transforms against painter-level output; a bug can live after frameAt.

6. Rerender the exact same sequence after each fix. Report the before/after scores for the same frame pair. Do not say a snap is fixed unless the same transition has been rerendered and inspected.

7. Keep only durable tests. Delete scratch diff tests/scripts before commit. If the bug was real, add a small regression test that asserts the measured failure mode directly, such as max visible center delta, no support re-anchor, monotonic camera movement, or bounded joint displacement.

Minimal Dart Diff Core

Use this core inside a throwaway Flutter test after rendering each frame to rawRgba bytes:

_VisualDiff _diff(
  Uint8List a,
  Uint8List b,
  int width,
  int height,
  int from,
  int to,
) {
  var changedPixels = 0;
  var totalDelta = 0;
  var minX = width;
  var minY = height;
  var maxX = 0;
  var maxY = 0;
  var sumX = 0.0;
  var sumY = 0.0;

  for (var y = 0; y < height; y++) {
    for (var x = 0; x < width; x++) {
      final offset = (y * width + x) * 4;
      final delta =
          (a[offset] - b[offset]).abs() +
          (a[offset + 1] - b[offset + 1]).abs() +
          (a[offset + 2] - b[offset + 2]).abs() +
          (a[offset + 3] - b[offset + 3]).abs();
      if (delta < 36) continue;
      changedPixels++;
      totalDelta += delta;
      minX = math.min(minX, x);
      minY = math.min(minY, y);
      maxX = math.max(maxX, x);
      maxY = math.max(maxY, y);
      sumX += x;
      sumY += y;
    }
  }

  final mean = changedPixels == 0 ? 0.0 : totalDelta / changedPixels;
  return _VisualDiff(
    from: from,
    to: to,
    changedPixels: changedPixels,
    meanChangedDelta: mean,
    score: changedPixels * mean,
    minX: changedPixels == 0 ? 0 : minX,
    minY: changedPixels == 0 ? 0 : minY,
    width: changedPixels == 0 ? 0 : maxX - minX + 1,
    height: changedPixels == 0 ? 0 : maxY - minY + 1,
    cx: changedPixels == 0 ? 0 : sumX / changedPixels,
    cy: changedPixels == 0 ? 0 : sumY / changedPixels,
  );
}

Reporting Format

Keep reports factual and phase-addressable:

Top temporal diffs, 240-frame dance ensemble:
1. 59->60 p=0.2458-0.2500 score=36.9M box=... centroid=...
   Read: f059.png, f060.png, diff_059_060.png
   Diagnosis: whole dancer re-anchors horizontally at support handoff.
2. ...

After patch:
59->60 score 36.9M -> 24.3M; inspected diff shows no whole-body teleport,
remaining delta is pose/silhouette. New worst: ...

Never collapse this to "looks fixed" without the numbers and image inspection.

Cleanup

• Scratch tests should use disposable names such as test/features/character/zzz_temporal_diff_test.dart.
• Scratch scripts can live under tool/, but remove them before commit unless the user explicitly asks to keep a maintained diagnostic.
• Diff PNGs belong in ignored build output.
• Commit only the production fix and focused regression tests.

See Also

• character-motion-review-panel for expert scoring after the discontinuities are measured and localized.
• choreo-phrase-authoring when the fix should become labelled move data rather than another anonymous key tweak.