AI: Animation in-betweening (ONNX)

[fernandotonon]

Parent epic: #397

Goal

Fill missing keyframes in sparse animations using Robust Motion In-betweening (Ubisoft, 2020) — a small transformer model that predicts intermediate poses given start/end keyframes and a target duration.

Why

Curve-editor and dope-sheet workflows often leave large gaps between keyframes. Hand-authoring smooth transitions is tedious. Local ML in-betweening is small enough to ship and produces noticeably better results than naive linear/Bezier interpolation.

Scope

• ONNX-exported model (~10 MB), distributed via ModelDownloader.
• Integrates into the Curve Editor (issue Curve editor: tangent handle drag + resample-into-track (D3 follow-up) #382) and the dope sheet (Phase 5 slice D1 — Dope sheet multi-select + copy/paste/bulk-move #374/Phase 5 slice D2 — Dope sheet per-channel rows #376).
• Action: select two adjacent keyframes → "In-between with N frames" → model predicts the gap.
• CLI: qtmesh anim model.fbx --in-between --gap-frames 30 -o filled.fbx.
• MCP: motion_in_between tool.
• Skeleton retargeting: the model is trained on a specific skeleton; document that it works best on humanoid skeletons close to the model's training distribution. Provide a clear fallback to spline interpolation when retargeting fails.

Acceptance Criteria

• Two human-pose keyframes plus a target duration produce smooth, plausible intermediate motion.
• Visibly better than linear interpolation on a test animation in the test fixtures.
• Falls back to spline interpolation with a warning when the input skeleton is incompatible.
• CLI/MCP/GUI parity.
• Sentry breadcrumb ai.assist.in_between recorded.

Dependencies

• ONNX Runtime (AI: PBR map synthesis from albedo (DeepBump-style, ONNX) #404).
• Curve Editor + dope sheet (already shipped).

Effort

~10 days.