rohm-onnx

RoHM ONNX evaluation bundle

This repository contains the mesh-free ONNX evaluation bundle used by the experimental RoHM post-processor in KitsuMate's Blender motion-capture extension.

Status

This is an experimental manual-evaluation release, not a validated StableMotion replacement. The current release decision is HOLD_STABLEMOTION: FP32 denoiser parity and the Windows WebGPU performance gate pass, but representative motion-quality evaluation and physical Linux/macOS runtime validation remain incomplete. See reports/gate_status.json for the exact completed and blocking gates.

Published profile

• Profile: fp32
• Download size: 222,030,103 bytes
• Upstream RoHM revision: 57ba22cde1595285276ddb88c8c82b25bf0d9722
• Runtime: NumPy and ONNX Runtime
• Intended providers: CPU, WebGPU, and CoreML
• Graphs: PoseNet and one gated TrajNet-Control graph; gate 0 reproduces vanilla TrajNet and gate 1 enables the controlled pass

The extension downloads only bundles/fp32/ and verifies every file against bundles/fp32/manifest.json.

Safety and dependency boundary

The published files contain no PyTorch checkpoints, pickle files, SMPL/SMPL-X assets, training data, or conversion environment. SMPL-X tensors embedded in the upstream checkpoints were removed before ONNX export, and the graphs were scanned for body-model initializer signatures. Production inference does not import or construct smplx; unavailable shape coefficients are treated as neutral and predicted shape output is ignored.

The small normalization data from upstream pickle files was read only inside the isolated conversion environment, deduplicated, and converted to metadata/normalizers.npz.

Precision selection

FP32 is the only published profile. FP16 was slower than FP32 in the tested CPU and WebGPU paths. Dynamic INT8 PoseNet produced about 32.289% relative L2 error, and dynamic INT8 TrajNet was slower than FP32. Calibrated static INT8 remains unavailable without representative motion calibration data.

License

RoHM's upstream license is reproduced in LICENSE.md and inside the downloadable bundle. Review it before redistribution or use. This repository does not include separately licensed SMPL/SMPL-X body-model files or benchmark datasets.

Upstream

• RoHM: https://github.com/sanweiliti/RoHM