RFD 1045: Model image for kimodo_text_to_motion
Problem
Kimodo turns a sentence into motion. It emits SOMA, and the client needs VRM humanoid tracks. The retarget between them is the hard part, and it is not the model.
The model is small at 0.6 GB in bf16. The packaging risk is the skeleton contract.
Decision
Return the SOMA result and the VRM result as separate outputs. The SOMA output is the model. The VRM output is a retarget, and a caller that has its own rig wants the SOMA.
Do not hide the retarget. A single VRM output makes the model look wrong when the retarget is what failed.
See DETAILS.md for the model’s memory, the predict() interface, and the motion validation gate this model image runs.