VeloMocap FlowHMR ONNX
motion capture model conversion
1 min readupdated
This repository provides a single-step flow-matching ONNX graph conversion of the FlowHMR model along with packed weights. It lets VeloMocap generate human motion from video without requiring PyTorch. Based on the original FlowHMR (arXiv:2610.03691), the weights are exported in fp32 without retraining. It is aimed at developers and researchers working on motion capture and integrating ONNX Runtime into Rust pipelines.
Key features
- ONNX graph and weights totaling around 1.7 GB for VeloMocap integration.
- RoPE tables and attention masks serve as graph inputs, leaving the Euler loop and decoding in Rust.
- Verified accuracy: maximum absolute difference against the PyTorch reference is 9.7e-3.
- Non-commercial license based on the upstream FlowHMR Model Weights license.
What you can do
- Drop the files into the models/flowhmr/ directory of a VeloMocap installation to run without PyTorch.
- Run the model with ONNX Runtime inside compatible motion pipelines.
Example
Place flowhmr_mmdit.onnx and flowhmr_mmdit.weights.bin into the models/flowhmr/ folder of your VeloMocap repository.
Read more
DetailsFiles: flowhmr_mmdit.onnx (3.9 MB), flowhmr_mmdit.weights.bin (1.7 GB).
Error rate: maximum absolute difference vs. PyTorch is 9.7e-3, mean is 2.1e-4.
License: Non-commercial (FlowHMR Model Weights License).










Comments