HMC - Human Motion Creator

An integrated platform for generating realistic 3D human motion from text descriptions, combining Tencent's HY-Motion 1.0 text-to-motion system with SMPLX and Meta's MHR (Momentum Human Rig). It is compatible with Meta Movement SDK high-fidelity full-body avatar, which means it is XR-READY.

Features

• Text-to-Motion Generation: Convert natural language descriptions into 3D human motion sequences using HY-Motion 1.0
• High-Fidelity Body Model: Meta MHR with identity (45 params), pose (204 params), and facial expression (72 params) control
• Format Conversion: Convert between SMPL, SMPLX, and MHR body model formats
• Multiple LOD Support: Generate meshes at 7 different levels of detail (LOD 0-6)
• Web Interface: Interactive Gradio UI for motion generation and visualization
• 3D Visualization: ScenePic-based interactive HTML visualizations

Supported FBX Templates

Template	Preview	Description
Wooden Model		Original wooden from HY-Motion-1.0
SMPLX		SMPLX body model
Meta Movement SDK		High-fidelity rig from Meta Movement SDK

Project Structure

HMC/
├── app.py                  # Main Gradio web interface
├── conf.py                 # Output path configuration
├── pixi.toml               # Pixi package manager config
├── HY-Motion-1.0/          # Text-to-Motion generation (submodule)
│   ├── hymotion/           # Core motion generation library
│   ├── ckpts/              # Model checkpoints
│   └── examples/           # Example prompts
├── MHR/                    # Momentum Human Rig (submodule)
│   ├── mhr/                # Core MHR package
│   └── tools/              # Conversion & visualization tools
├── scripts/                # Conversion and visualization scripts
│   ├── simple_convert.py   # SMPLX/FBX to MHR conversion
│   └── visualize_output.py # 3D mesh visualization
├── assets/                 # Model assets and data files
│   ├── SMPLX_NEUTRAL.npz   # SMPLX body model
│   ├── mhr_model.pt        # MHR TorchScript model
│   ├── lod*.fbx            # FBX rigged models (LOD 0-6)
│   └── *mapping.npz        # Joint mapping files
└── output/                 # Generated outputs (created at runtime)

Requirements

• Python 3.12+
• CUDA 13.0+
• PyTorch 2.8.0+

This project is developed on Ubuntu 24.04 with one RTX 5090 GPU.

Installation

# Clone with submodules
git clone --recursive https://github.com/your-repo/HMC.git
cd HMC

# Download and unzip the model assets
curl -OL https://github.com/luffy-yu/HMC/releases/download/v1.0.0/assets.zip
unzip assets.zip -d assets/

# Install dependencies
pixi install

# Activate environment
pixi shell

Download Model

Download the model checkpoints from HY-Motion-1.0 ckpts README and place them in the following directories:

Model	Path
HY-Motion	./HY-Motion-1.0/ckpts/tencent/HY-Motion-1.0/HY-Motion-1.0-Lite
Prompt Engineering	./HY-Motion-1.0/ckpts/Text2MotionPrompter

Text2MotionPrompter would be unnecessary if prompt engineering is disabled.

Usage

Web Interface

Launch the Gradio web interface:

python app.py

# Disable prompt engineering (required even for RTX 5090)
python app.py --disable-prompt-engineering

Access the interface at http://localhost:7860

Command-line options:

Option	Description	Default
--host	Server host address	localhost
--port	Server port number	7860
--disable-prompt-engineering	Disable text rewriting and duration estimation	False

--disable-prompt-engineering is even needed for RTX 5090 GPU.

Examples:

python app.py --host 0.0.0.0 --port 7860   # Run on all interfaces, port 7860
python app.py --disable-prompt-engineering  # Disable prompt engineering
python app.py --help                        # Show all options

Command Line

Generate motion from text:

cd HY-Motion-1.0
python local_infer.py --prompt "A person walks forward slowly"

Convert SMPLX to MHR:

# Default template (no fbx output)
python scripts/simple_convert.py --smplx assets/SMPLX_NEUTRAL.npz -i assets/tpose_poses.npy -o output/default_template

# SMPLX template (copy animation from --smplx to lod1.fbx and output walking.fbx to output/smplx_template)
python scripts/simple_convert.py --smplx assets/walking.fbx --fbx-template assets/lod1.fbx -o output/smplx_template

# Meta Movement SDK Avatar template (copy animation from --smplx to high_fidelity_rig.fbx and output walking.fbx to output/smplx_template)
python scripts/simple_convert.py --smplx assets/walking.fbx --fbx-template ./assets/high_fidelity_rig.fbx -o output/hifi_template

high_fidelity_rig.fbx is from Meta Movement 72.0.0 (com.meta.movement, https://github.com/oculus-samples/Unity-Movement.git, Assets\Samples\Meta Movement\72.0.0\Advanced Samples\HipPinning\Models\high_fidelity_rig.fbx).

Visualize output:

# Use default output name
python scripts/visualize_output.py --output_dir output/default_template

# Customize the html file name
python scripts/visualize_output.py --output_dir output/smplx_template --html_name my_visualization.html

# Loop folders to generate
python scripts/visualize_output.py --output_dir output/

Output Formats

• Motion Sequences: .npy, .fbx
• 3D Meshes: .ply, .fbx
• Visualizations: Interactive HTML files

Acknowledgments

• Tencent HY-Motion 1.0 - Text-to-Motion generation
• Meta MHR - Momentum Human Rig body model
• SMPL-X - Expressive body model

License

See individual submodule licenses for HY-Motion and MHR components.