Summary

HY-Motion 1.0 generates 3D human motion from text prompts, integrating into 3D animation pipelines. It scales Diffusion Transformer (DiT) and Flow Matching models to billion-parameter levels, delivering state-of-the-art instruction-following and motion quality, surpassing existing open-source alternatives.

How It Works

This project utilizes Diffusion Transformer (DiT) and Flow Matching architectures for text-to-3D motion generation. Its key innovation is scaling these models to billion-parameter size, significantly improving instruction understanding and motion fidelity. Training involves three stages: extensive large-scale pre-training, high-quality fine-tuning, and reinforcement learning refinement for enhanced naturalness and instruction adherence.

Quick Start & Requirements

Installation requires cloning the repo, installing PyTorch, ensuring git-lfs is present (git lfs pull), and running pip install -r requirements.txt. The 1.0B parameter model needs 26GB VRAM minimum; the 0.46B Lite version requires 24GB. Lower VRAM usage is possible via --num_seeds=1, prompts <30 words, and motion length <5 seconds. Prompt engineering features can be disabled (DISABLE_PROMPT_ENGINEERING=True) to reduce VRAM. Local inference scripts and a Gradio app are available.

Highlighted Details

• Achieves state-of-the-art performance in instruction-following and motion quality.
• Pioneers billion-parameter scale for DiT-based text-to-motion models.
• Employs advanced three-stage training: pre-training, fine-tuning, and reinforcement learning.

Maintenance & Community

The README acknowledges contributions from numerous open-source projects. No specific community channels, roadmaps, or direct contributor details beyond library acknowledgements are provided.

Licensing & Compatibility

No explicit software license is stated in the README. While the project utilizes and acknowledges many open-source libraries, the specific terms for HY-Motion 1.0's use, modification, and distribution are undefined, potentially impacting commercial adoption.

Limitations & Caveats

Unsupported features include non-humanoid characters, subjective visual attributes (emotions, clothing), environment/camera details, and multi-person interactions. Special modes like seamless loops are also excluded. English prompts are recommended; other languages require a separate prompter module.