Hy-Motion 1.0: Scaling Flow Matching Models for 3D Motion Generation (github.com)

Tencent has unveiled HY-Motion 1.0, an innovative series of text-to-3D human motion generation models employing Diffusion Transformer (DiT) and Flow Matching techniques. This powerful model allows users to create skeleton-based 3D animations from straightforward text prompts, streamlining integration into 3D animation workflows. Notably, HY-Motion 1.0 is the first to scale DiT-based models to a billion-parameter size, resulting in a marked enhancement in instruction-following abilities and overall motion quality when compared to existing open-source alternatives. The significance of HY-Motion 1.0 lies in its state-of-the-art performance achieved through a rigorous three-stage training process. It involves initial pre-training on over 3,000 hours of diverse motion data, followed by fine-tuning on 400 hours of high-quality 3D motion data, and reinforcement learning to improve naturalness and instruction adherence. With robust support for macOS, Windows, and Linux, developers can easily access and implement this advanced capability through available pretrained models and inference code on HuggingFace, marking a significant advancement in the realm of AI-driven motion generation.