Tencent-Hunyuan/HY-Motion-1.0

What it solves

HY-Motion 1.0 addresses the challenge of generating high-quality, natural 3D human animations from text descriptions. It improves upon previous open-source models by enhancing the semantic alignment between the text prompt and the resulting motion, as well as the overall smoothness and realism of the character's movements.

How it works

The project utilizes a Diffusion Transformer (DiT) architecture combined with Flow Matching. It scales these models to the billion-parameter level to improve instruction-following capabilities. The training process consists of three stages:

1. Large-Scale Pre-training: Learning a broad motion prior from over 3,000 hours of motion data.
2. High-Quality Fine-tuning: Refining motion detail and smoothness using 400 hours of curated data.
3. Reinforcement Learning: Using human feedback and reward models to further improve motion naturalness and prompt adherence.

Who it’s for

This tool is designed for developers and 3D animators who need to generate skeleton-based 3D character animations that can be integrated into existing 3D animation pipelines.

Highlights

• Billion-Scale Parameters: The first DiT-based text-to-motion model to reach the billion-parameter scale for better instruction understanding.
• Three-Stage Training: A comprehensive pipeline involving pre-training, high-quality fine-tuning, and RLHF.
• Hunyuan-1.0 and Lite versions: Offers both a standard 1.0B parameter model and a lightweight 0.46B version.
• VLM-based Evaluation: Includes SSAE (Structured Semantic Alignment Evaluation) to assess how well generated motions align with text prompts.