HY-Motion 1.0 - Tencent Hunyuan's open-source large-scale text-to-3D motion generation model

What is HY-Motion 1.0?

HY-Motion 1.0 (Hunyuan Motion 1.0) is a billion-parameter text-to-3D motion generation model open-sourced by Tencent's Hunyuan team. Based on the Diffusion Transformer (DiT) architecture and flow matching mechanism, it can generate high-fidelity, smooth, and diverse 3D character skeletal animations with just one natural language description. The model employs a full-stage training strategy, including pre-training on over 3000 hours of diverse motion data, fine-tuning on 400 hours of carefully selected high-quality data, and optimization through reinforcement learning combined with human feedback and reward models. It covers over 200 actions across 6 major categories, outputs in SMPL-H skeletal format, and supports atomic actions, combined sequences, and concurrent action generation. In performance evaluation, instruction compliance reaches 78.6%, and the average motion quality score is 3.43 (out of 5), comprehensively surpassing open-source baselines such as MoMask and DART.

Main functions of HY-Motion 1.0

• Text-driven action generationIt generates high-quality 3D skeletal animations directly from natural language descriptions, achieving efficient conversion from text to motion.
• Diverse motion coverageIt covers more than 200 actions in 6 major categories, including basic movement, sports competition, social leisure, and game character actions, to meet the needs of various scenarios.
• High-quality motion outputIt supports the SMPL-H bone format, generating smooth and natural animations with rich motion details, making it suitable for high-requirement animation production.
• Compatible with mainstream toolsIt is directly compatible with mainstream 3D tools such as Blender, Unity, and Unreal Engine, making it convenient for users to quickly integrate and use.
• Flexible output optionsIt supports the generation of atomic actions, combined sequences, and concurrent actions to meet animation needs of varying complexity.
• Open source and ease of useIt provides complete inference code, pre-trained model weights, and detailed documentation, supports multiple operating systems, and is easy to use.

Technical principles of HY-Motion 1.0

• Based on Diffusion Transformer architectureIt adopts the advanced Diffusion Transformer (DiT) architecture and combines it with a flow matching mechanism to efficiently transform text descriptions into high-quality 3D skeletal animations.
• Full-stage training strategyThe model optimizes its performance through three stages of training: pre-training on over 3,000 hours of diverse action data to learn a wide range of action patterns; fine-tuning on 400 hours of carefully selected high-quality data to improve action accuracy; and finally, optimization through reinforcement learning combined with human feedback and reward models to ensure the naturalness and accuracy of generated actions.
• Stream matching mechanismBy utilizing flow matching technology, the model can better capture the continuity and dynamic changes in motion data, generating smoother and more natural animations.
• Multimodal fusionThe text modality and action modality are deeply integrated. The text encoder converts the text description into a feature vector, which is then combined with the action generator to achieve accurate action generation.
• Reinforcement learning optimizationBy introducing a reinforcement learning mechanism and combining human feedback and reward models, the generated actions are evaluated and optimized, further improving the model's action generation quality and instruction compliance ability.

Project address for HY-Motion 1.0

• Project official websitehttps://hunyuan.tencent.com/motion
• Github repository: https://github.com/Tencent-Hunyuan/HY-Motion-1.0
• Huggingface model libraryhttps://huggingface.co/tencent/HY-Motion-1.0
• arXiv technical paper: https://arxiv.org/pdf/2512.23464

Application scenarios of HY-Motion 1.0

• Film and animation productionQuickly generate character animations, reduce animation production costs and time, and improve creative efficiency.
• Game developmentProvide diverse actions for game characters, enhance the gaming experience, and adapt to different game scenarios.
• Virtual anchors and digital humansGenerate natural and fluid movements, enhancing the interactivity and realism of virtual avatars.
• Education and TrainingCreate instructional animations to simulate complex action processes and assist in teaching and training.
• Advertising and MarketingCreate personalized animated ads to attract viewers' attention and improve marketing effectiveness.
• VR/AR applicationsGenerate real-time motion for virtual reality and augmented reality content to enhance immersion.