Hunyuan Motion 1.0, a text to human motion model

The Dawn of Billion-Parameter Motion Synthesis

The landscape of Generative AI has been shifting at a tectonic pace. We have witnessed the explosion of text-to-image models that can conjure surrealist art in seconds, and text-to-video models that are beginning to blur the lines between reality and simulation. However, a critical piece of the puzzle has often remained elusive or stuck in the uncanny valley: 3D human motion generation.

For developers, game designers, and animators, creating realistic 3D character movements is a labor-intensive process involving motion capture suits, expensive studios, and hours of manual keyframing. Enter HY-Motion 1.0 (also known as Hunyuan Motion 1.0), a groundbreaking release that promises to democratize high-quality animation.

HY-Motion 1.0 represents a significant leap forward. It is not just another incremental update; it is the first successful attempt to scale a Diffusion Transformer (DiT) based flow matching model to the billion-parameter scale within the motion generation domain. But what does that actually mean for the future of AI and 3D content? Let’s explore.

Who is Behind HY-Motion 1.0?

This sophisticated model is the brainchild of the Tencent Hunyuan 3D Digital Human Team. Tencent, a global technology giant known for its dominance in the gaming and social media sectors (with titles like Honor of Kings and PUBG Mobile under its umbrella), has a vested interest in perfecting digital avatars.

The release of HY-Motion 1.0 aligns with Tencent’s broader strategy to lead in the embodied intelligence and digital content creation sectors. By open-sourcing this tool, they are not only showcasing their technical prowess but also accelerating the industry’s transition from manual animation to AI-assisted generation. The project involves a massive collaboration of researchers, engineers, and artists, aiming to solve the data bottleneck that has historically plagued motion generation research.

What Exactly is HY-Motion 1.0?

At its core, HY-Motion 1.0 is a text-to-motion generation model. You feed it a text prompt and it outputs a 3D skeleton animation sequence that can be used in game engines or 3D software.

However, the architecture under the hood is what makes it special. Unlike previous models that relied on smaller, standard diffusion processes, HY-Motion 1.0 utilizes a Diffusion Transformer (DiT) architecture combined with a Flow Matching objective.

The Power of Flow Matching

Traditional diffusion models generate data by gradually removing noise. Flow Matching is a more recent and efficient paradigm. It constructs a continuous probability path (an Ordinary Differential Equation or ODE) that bridges the gap between random noise and complex motion data. In simple terms, it finds the most optimal path to transform noise into a realistic movement, resulting in faster generation and higher stability during training.

The Billion-Parameter Scale

Size matters in AI. We have seen Scaling Laws apply to Large Language Models (LLMs) like GPT-5, where more parameters generally equal better reasoning and understanding. HY-Motion 1.0 applies this logic to motion. By scaling the model to over 1 billion parameters, the system gains a superior understanding of complex textual instructions and can generate more nuanced, physically plausible movements compared to its smaller predecessors like MoMask or DART.

What Makes HY-Motion 1.0 Different? The Scale-Then-Refine Strategy

The true innovation of HY-Motion 1.0 lies in its training methodology. The team at Tencent didn’t just throw data at a neural network; they engineered a comprehensive, three-stage training paradigm that mimics how humans learn skills: first broadly, then precisely, and finally, through feedback.

Stage 1: Large-Scale Pretraining (The Learning to Move Phase)

The model was first trained on a massive dataset comprising over 3,000 hours of motion data. This dataset is a hybrid of high-quality motion capture and in-the-wild video data processed from the HunyuanVideo dataset. Using tools like GVHMR, they extracted 3D human tracks from millions of video clips.

In this phase, the model learns the grammar of human movement. It learns what a walk looks like, how joints bend, and the general correlation between text and action. However, because some of this data comes from raw videos, the output can sometimes be jittery or contain artifacts (like foot sliding).

Stage 2: High-Quality Fine-Tuning (The “Refinement” Phase)

To fix the jitter and increase fidelity, the model undergoes fine-tuning on a curated subset of 400 hours of high-quality data. This data is rigorously filtered to remove abnormalities and is paired with manually corrected textual descriptions. This stage sharpens the model, significantly reducing artifacts and ensuring the motion looks professional and smooth.

Stage 3: Reinforcement Learning (The Human Alignment Phase)

This is where HY-Motion 1.0 truly distinguishes itself from the competition. Most motion models stop at supervised learning. Tencent went a step further by implementing Reinforcement Learning (RL) using:

• DPO (Direct Preference Optimization): The model is trained on pairs of motions where human judges picked a winner based on quality and instruction adherence.
• Flow-GRPO: A specialized RL technique for flow matching models that optimizes for specific rewards, such as physical plausibility (penalizing foot sliding) and semantic consistency.

This final stage ensures that the generated motions aren’t just statistically likely, but actually look good to human eyes and strictly follow the user’s text prompts.

Technical Deep Dive: The Architecture

For the technical enthusiasts reading artificial-intelligence.be, the architecture of the HY-Motion DiT is fascinating. It employs a hybrid Transformer design similar to state-of-the-art video models.

Dual-Stream and Single-Stream Blocks

The network processes text and motion in two ways. First, Dual-Stream Blocks process motion latents and text tokens independently but allow them to interact via cross-attention. This prevents the “noise” from the motion generation process from corrupting the semantic meaning of the text. Later, Single-Stream Blocks concatenate the two modalities for deep fusion.

Advanced Text Encoders

To understand prompts, the model uses a hierarchical strategy. It employs Qwen3-8B (a powerful LLM) for fine-grained detail and CLIP-L for global semantic understanding. A “Bidirectional Token Refiner” is used to adapt the causal (left-to-right) nature of the LLM into a bidirectional context suitable for motion generation.

The SMPL-H Standard

The output is standardized on the SMPL-H skeleton (22 joints). This is a crucial detail for developers, as SMPL-H is a widely accepted format in the research and animation industry, making it easier to retarget these motions onto custom 3D characters in engines like Unity or Unreal Engine.

How to Use HY-Motion 1.0

One of the most exciting aspects of this release is that it is open-source. The weights and code are available on Hugging Face, allowing you to run this on your own hardware (provided you have a decent GPU).

1. Installation and Requirements

The model is built on PyTorch. To get started, you will need to clone the repository from GitHub or Hugging Face. The installation is standard for Python-based AI projects:

• Install PyTorch (ensure CUDA support is enabled).
• Install dependencies listed in the `requirements.txt` (includes libraries like `diffusers`, `smplx`, and `transformers`).
• Download the pre-trained weights (check the `ckpts/README.md` in their repository).

2. Running Inference

Tencent provides a Command Line Interface (CLI) script for batch processing. This is useful if you want to generate hundreds of motions overnight. However, for exploration, they also provide a Gradio App. By running the Gradio script, you launch a local web interface (usually at `http://localhost:7860`) where you can type prompts and see the rendered skeleton in real-time.

3. Hardware Considerations

Since this is a billion-parameter model, VRAM is a consideration. The developers note that to reduce GPU VRAM requirements, you should limit the batch size (seeds) to 1 and keep motion length under 5 seconds. There is also a “Lite” version (0.46B parameters) available if you are running on more modest hardware, though the 1B version offers superior instruction following.

4. Prompting Best Practices

To get the best results, treat the model like a junior animator who needs clear instructions:

• Language: Use English. The model is optimized for English prompts.
• Length: Keep prompts under 60 words.
• Specificity: Focus on limb and torso movements. Instead of “he is happy,” try “a person jumping repeatedly with arms raised in joy.”
• Structure: The system includes a Prompt Rewrite module powered by an LLM that automatically optimizes your input, but starting with a clear action description helps.

Performance vs. Competitors

In the fast-moving world of AI, benchmarks are everything. HY-Motion 1.0 has been tested against leading open-source models like MoMask, DART, and GoToZero.

Using a metric called SSAE (Structured Semantic Alignment Evaluation), which uses Video-VLMs to judge if the motion matches the text, HY-Motion 1.0 scored 78.6%. In comparison, MoMask scored 58.0% and DART scored 42.7%. This is a massive margin, highlighting the effectiveness of the scaling laws and the RLHF (Reinforcement Learning from Human Feedback) stage.

Qualitatively, users report significantly less foot sliding, common glitch where a character’s feet appear to glide over the floor rather than planting firmly. The physics compliance reward in the training phase specifically targeted this issue.

Limitations and Future Outlook

While HY-Motion 1.0 is a state-of-the-art tool, it is not magic. The developers have been transparent about its current limitations:

• Complex Instructions: While it beats competitors, it can still struggle with highly nuanced or multi-stage instructions that require long-term memory.
• Human-Object Interaction (HOI): The model generates body movements, but it doesn’t understand external objects. If you ask for a person holding a cup, the hand will move as if holding a cup, but the fingers might not be perfectly positioned for the geometry of a specific object.
• Facial Expressions and Hands: The current version focuses on the 22 main body joints (SMPL-H without hand articulation details or facial blendshapes).

Despite these limitations, the release of HY-Motion 1.0 is a watershed moment. It proves that the Scaling Laws that powered the LLM revolution are equally applicable to 3D motion. By combining massive datasets, transformer architectures, and reinforcement learning, Tencent has set a new baseline for what is possible.

Conclusion

HY-Motion 1.0 is more than just a research paper; it is a practical, powerful tool that is now available to the world. For the community, this represents an exciting opportunity to experiment with the next generation of generative media. Whether you are looking to populate a digital twin of a city, create dynamic NPCs for an indie game, or simply explore the frontiers of AI movement, HY-Motion 1.0 provides the most robust foundation we have seen to date.

As we look toward 2026 and beyond, we can expect this scale-then-refine approach to be applied to even more complex tasks, eventually solving the interaction problems and bringing us closer to fully generative, interactive 3D worlds.

This article is based on the dutch article about Hunyuan Motion 1.0 on artificial-intelligence.be.