StableMotion ONNX
ONNX conversion for the StableMotion Correction sequence post-processor in Vision To Pose.
This model is published only for non-commercial research use. The StableMotion source code is MIT-licensed, while the checkpoint was trained on BrokenAMASS derived from AMASS and remains subject to applicable StableMotion, AMASS, SMPL, and source-dataset terms. This conversion grants no commercial rights and does not replace or expand any upstream license.
Limitations
- Intended only for non-commercial research and the Vision To Pose cleanup workflow.
- Trained at 20 FPS on BrokenAMASS; motion outside that domain may be missed or changed undesirably.
- StableMotion is probabilistic. The integration fixes seed
10, validates rotations, preserves samples outside detected intervals, and rolls back failed bakes, but cannot guarantee improvement for every animation. - Low-frame-rate input is upsampled and contains less recoverable temporal detail.
- Face, fingers, custom controls, and FK/IK switches are not model outputs. Results still depend on the source detector, target rig, and mapping quality.
Source and contract
- Upstream: https://github.com/Murrol/StableMotion
- Pinned source commit:
45d8836ce5cd7ae70f0065fdac672debfdfd6066 - Official EMA checkpoint member:
stablemotion/ema001000000.pt - Checkpoint archive SHA-256:
b117ea6f3cefd92d3bc804b311bdcb003c4a4c6382bb697d10d7629c6665053c - Model clock: 20 FPS
- Window: 100 frames
- Representation: 232 motion features plus one quality-label feature
- Sampling: official 50-step cosine diffusion, basic detection, threshold
0.5, one-frame mask dilation, then fixing/inpainting
The immutable runtime bundle is flat under onnx/. onnx/manifest.json declares every file size and SHA-256, graph names, source checkpoint identity, and the self-test contract.
Validation status
Technical conversion gates pass locally:
- PyTorch EMA-to-ONNX denoiser parity
- Full detect/dilate/fix parity with identical masks
- 10, 20, 24, 25, 29.97, 30, 50, 59.94, and 60 FPS plus variable-rate timing
- Blender quaternion continuity and FK/IK safety tests
- Recorded MediaPipe and SAM3D Body VRoid pose preservation
- Injected 155-degree arm discontinuity reduced to approximately 3.46 degrees without a new flip
onnx/parity_report.json contains the detailed measurements. Use of the model remains subject to the non-commercial and upstream-license limitations above.