AI: Text-to-motion via MDM (ONNX) — research/ambitious

[fernandotonon]

Parent epic: #397

Goal

Generate skeletal animation from a text prompt using MDM (Motion Diffusion Model) or a similar text-conditioned motion model, retargeted to the current skeleton.

Status

Research-grade / highest risk in the epic. Tier 4. Listed for transparency but should not be picked up until at least #409 (motion in-betweening) ships, since it shares much of the same retargeting plumbing.

Scope (sketch)

• ONNX export of MDM or a smaller successor (T2M-GPT, MotionGPT). Verify export feasibility before committing.
• Motion output is in SMPL format; needs a retargeting layer onto QtMeshEditor's Ogre skeletons.
• CLI: qtmesh anim model.fbx --generate "walking confidently" --duration 4 -o out.fbx.
• GUI: Animation Mode action "Generate from text".
• MCP: generate_motion tool.
• Documented as an experimental feature behind --experimental flag.

Open questions

• Which MDM variant produces the best ONNX-exportable model with acceptable quality?
• What is the realistic motion length / quality on CPU-only machines?
• Is there a lighter alternative (T2M-GPT) that runs comfortably on a MacBook?
• Retargeting strategy: pose-space remap vs IK-based?

Acceptance Criteria

• Spike: confirm ONNX export of at least one current motion-generation model.
• Prototype: text → motion → retargeted onto a test humanoid in QtMeshEditor.
• Decision: ship behind --experimental, postpone, or close as not feasible.

Effort

~3+ weeks if pursued, with significant uncertainty.