This project addresses the challenge of zero-shot text-to-motion generation by introducing MotionMillion, the largest human motion dataset (2M+ sequences), and MotionMillion-Eval, a comprehensive benchmark. It targets researchers in computer vision, graphics, and robotics, offering a scalable model (up to 7B parameters) with strong generalization capabilities for versatile motion synthesis.

How It Works

The approach uses a scalable architecture trained on the extensive MotionMillion dataset, incorporating LLAMA3.1-8B for prompt rewriting. It introduces MotionMillion-Eval, a benchmark designed to rigorously assess zero-shot motion generation performance, particularly for out-of-domain and complex compositional motions.

Quick Start & Requirements

Setup requires Conda with Python 3.8.11 and PyTorch 2.4.1. Install dependencies via pip install -r requirements.txt and conda install conda-forge::git-lfs. Essential external downloads include SMPL+H/DMPL body models (to ./body_models/), human model files (Google Drive), and materials via bash prepare/download_glove.sh, bash prepare/download_t2m_evaluators_on_motionmillion.sh, and bash prepare/download_T5-XL.sh. Pretrained 3B/7B models are available via bash prepare/download_pretrained_models.sh. Dataset preparation details are "Comming Soon!". Inference and training scripts are provided.

Highlighted Details

• Introduces MotionMillion, a dataset exceeding 2,000 hours and 2 million high-quality human motion sequences.
• Presents MotionMillion-Eval, a comprehensive benchmark for zero-shot motion generation evaluation.
• Features a scalable model architecture validated up to 7 billion parameters.
• Achieved an ICCV 2025 Highlight award.
• Demonstrates strong generalization to out-of-domain and complex compositional motions.

Maintenance & Community

The project is affiliated with multiple universities and Shanghai AI Laboratory. Contact information for corresponding authors is provided. No explicit community channels (e.g., Discord, Slack) or roadmap links are detailed.

Licensing & Compatibility

Licensed under the Apache License, which generally permits commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

Dataset preparation details are marked "Comming Soon!", suggesting potential incompleteness. Setup involves downloading and organizing large external models and running multiple bash scripts, presenting a complex initial hurdle. The project is a recent development associated with ICCV 2025.