Serqan/MotionSmith-Mocap-Runtime

MotionSmith Mocap Runtime

Model-free Windows x64 runtime for the MotionSmith Mocap Blender extension. This repository distributes software, not a trained model.

Download

MotionSmith Mocap Runtime 1.0.0 — Windows x64

Size: 3,141,367,826 bytes (approximately 3.14 GB / 2.93 GiB).

SHA-256: c764ffb8b9f4b1dade5c22cb811a6629da2ee2d5f8b5ad500d4d42424c499b8c

See runtime-manifest.json for exact package versions and validation results, and SHA256SUMS.txt for the archive checksum.

Included

• Application-local Python 3.12.10, Windows x64; no system Python installation required.
• PyTorch 2.11.0 with CUDA 12.8 and validated GEM-X/SOMA-X Python dependencies.
• MotionSmith Mocap worker and shared MotionSmith GPU reservation.
• YOLOX COCO capture configuration and applicable third-party licence/notice files.

Model weights, SOMA geometry assets, FFmpeg binaries, Blender, footage and user projects are not included. The weights directory contains a configuration file only.

Installation

1. Extract the entire MotionSmithMocap folder into a writable folder outside AppData. Keep its directory structure.
2. In MotionSmith Mocap preferences, set Runtime Root and Backend folder to the extracted MotionSmithMocap folder. Choose a writable output folder.
3. Install model assets separately under weights/ and soma/ using the official providers below. The archive never downloads models automatically.
4. Install FFmpeg separately and ensure ffmpeg is available on PATH. Use a CUDA 12.8-compatible NVIDIA driver and the Microsoft Visual C++ x64 runtime.

Model assets obtained separately

• NVIDIA GEM-X: body_s240000.ckpt, body_config.json, vitpose.pth, motion_native77_e774_s292175.ckpt, motion_stats_raw.npz into weights/.
• NVIDIA SOMA-X: native SOMA v0.3.3 assets into soma/.
• Official Megvii YOLOX COCO detector: yolox_x.onnx into weights/. The default configuration does not use the non-commercial Human-Art detector.

After these prerequisites are installed, capture runs locally/offline in a separate worker process. Heavy model imports do not run inside Blender.

Validation

The sealed archive was extracted to a different folder containing spaces. Its Python and core packages imported independently of system Python, with deliberately invalid PYTHONHOME/PYTHONPATH overrides. CUDA tensor execution passed. Using separately supplied existing model assets only in the extracted test folder, the worker completed a real 60-frame/30-fps GEM -> Contact/IK -> SOMA77 BVH capture and shut down normally. BVH readback verified 78 joints, finite transforms and proper rotations.

Validation was performed on Windows with an RTX 5070 Ti and the required external FFmpeg/driver prerequisites; it is not a fresh-machine or every-GPU certification. No test model assets are present in the uploaded archive.

Licensing

This is a mixed-component software distribution. MotionSmith-owned worker code is Copyright 2026 MotionSmith AI, all rights reserved. Third-party packages and NVIDIA-derived skeleton reference data retain their own licences. See LICENSES/README.md, the included component notices/package metadata, and runtime/Scripts/LICENSE.txt inside the archive. The runtime archive grants no licence to separately obtained model weights; their respective provider terms apply.