What is HY-Motion 1.0? A Deep Dive into Tencent’s Open-Source Model That Generates 3D Human Motion from Text (Features, How to Use, Required GPU, License Caveats)
- HY-Motion 1.0 is a family of models from Tencent’s Hunyuan team that can generate bone/skeleton-based 3D human motion from text instructions.
- The technical core is described as Diffusion Transformer (DiT) plus Flow Matching, scaled up to the billion-parameter class.
- A standard model (1.0B) and a lightweight model (0.46B) are released, with minimum VRAM guidance of 26GB / 24GB respectively (with tips to reduce usage via settings).
- Before using it, you must check the license: the license does not apply in the EU, UK, and South Korea, and for sufficiently large services (large MAU), you must apply for an additional license.
HY-Motion 1.0 is an AI model aimed at pulling the most time-consuming part of 3D animation production—creating motion—closer to you via text input. The “motion” here is not the final rendered video, but animation data in the form of skeleton (rig) motion that can be applied to a 3D character. In a world where hand-keying and motion capture dominate, you can think of HY-Motion 1.0 as an entrance that lets you start from language.
Based on publicly available primary sources (GitHub, Hugging Face, arXiv, the license text, and domestic reporting), this article summarizes, in a practical and “hands-on” way: what HY-Motion 1.0 can do, how it works, what compute it needs, how to start using it, and what to watch out for in commercial use and redistribution. I also include prompt examples and a sample adoption workflow so you can prepare to actually try it after reading.
What HY-Motion 1.0 Can Do: Text → 3D Human Motion (Skeleton) Generation
The core function of HY-Motion 1.0 is generating 3D human motion as skeleton-based animation from natural-language prompts. Official descriptions say the outputs can be “integrated into a variety of 3D animation pipelines,” suggesting it can serve as an entry point for workflows across games, film/animation, VR/AR, and research.
The paper abstract claims the model is scaled up for text-to-motion and strengthened for instruction-following, and it also notes coverage of 200+ motion categories.
Domestic reporting similarly cites examples ranging from basic actions like sitting, running, and jumping to sports motions, dance, and Tai Chi, emphasizing naturalness and fewer failures.
What matters here is that HY-Motion 1.0 is not a magical tool that produces a finished film shot. It’s an AI that generates motion assets for character animation. To reach a finished result, you still need to apply the motion to your rig, often retarget it, and then handle staging, camera work, lighting, editing, and so on. But if the “quickly get motion” step becomes shorter, you can iterate more, improving planning and direction because you’ve freed time for higher-level creative decisions.
Who Will Benefit Most? Concrete Target Users (Production/Development/Research Perspectives)
HY-Motion 1.0 tends to help most in the following cases. I’ll be a bit detailed here because adoption decisions often hinge less on flashy features and more on “which part of my work becomes lighter.”
1) Game developers (especially teams that prototype frequently)
In game production, reaching agreement on “how it should move” across planners, animators, and engineers can be a major effort. For example, before you hand-craft an action like “draw a sword, step in, and slash,” you often need a rough motion sketch first. HY-Motion 1.0 may let you start from text for that initial draft. Domestic reporting also points to game-relevant motions such as sword-and-shield actions.
2) Film/VFX teams who want to iterate previs and blocking faster
In previs, you often want the intention of movement and composition more than final polish. If you can generate motion direction from text and begin directorial discussion earlier, you reduce confusion later. Because HY-Motion 1.0 is geared toward generating reusable motion materials, it can fit well between storyboards and 3D blocking.
3) Solo creators / indie teams who lack time and manpower
In solo production, “work volume” often defeats ambition. A model like HY-Motion 1.0 can shorten the entry to motion creation and increase your “try and discard” cycles before hand-tweaking. More iteration frequently leads to better quality, so this can be genuinely helpful.
4) Research and education (motion generation, human motion understanding, data augmentation, etc.)
The project’s research position is explicit, and the training data pipeline is described in a structured way (pretraining on 3,000+ hours, fine-tuning on 400 hours of high-quality data, plus reinforcement learning with human feedback). For research/education, the fact that you can read “how it was made” is itself valuable.
Key Mechanism: DiT + Flow Matching, and Multi-Stage Training for “Following Instructions”
HY-Motion 1.0 is described as a text-to-motion generation model family built on Diffusion Transformer (DiT) and Flow Matching. The headline claim is scaling a DiT-based text-to-motion model to the ~1B parameter level to improve instruction-following and motion quality.
The training paradigm is presented in stages. Combining GitHub and domestic explanations yields the following pipeline.
- Large-scale pretraining on 3,000+ hours of diverse motion data (learning a broad prior over motion)
- Fine-tuning on 400 hours of high-quality 3D motion data (improving smoothness and details)
- Reinforcement learning using human feedback and/or a reward model (further aligning instruction understanding and naturalness)
Domestic reporting also mentions that, to convert vague prompts into structured instructions that the model can follow more reliably, the pipeline may incorporate an LLM (examples given include Gemini 2.5 Pro and Qwen, etc.). This reads as an effort to improve the completeness of the overall generation pipeline rather than relying on the motion model alone.
Model Lineup and Required GPU: Standard vs Lite, VRAM Guidance and “How to Make It Lighter”
At least the following two models are explicitly listed.
- HY-Motion-1.0 (Standard): 1.0B parameters, minimum VRAM guideline 26GB
- HY-Motion-1.0-Lite: 0.46B parameters, minimum VRAM guideline 24GB
This VRAM requirement is not “casual gaming GPU on one card” territory; it leans more toward production/research setups. However, the Hugging Face guidance includes tips to reduce VRAM usage, such as “set seed count to 1,” “keep prompts under 30 words,” and “keep motion length under 5 seconds.” If you generate many short motions and iterate, you may find a practical balance.
The GitHub instructions also provide an environment variable to disable an LLM-based “prompt engineering” feature if the Gradio app hits VRAM errors. In other words, you can prioritize the core model by turning off auxiliary features to make your environment more likely to work.
Supported OSes are described as macOS, Windows, and Linux, which helps in mixed production environments.
How to Start: Two Paths—Local Execution (CLI) and Demo (Space)
HY-Motion 1.0 provides inference code and pretrained weights, with a fairly standard flow: clone repo → install dependencies → place weights → run inference scripts.
Here’s a sample “feel of the steps” based on primary sources (details will vary by environment, so following the official README structure is safest).
git clone https://github.com/Tencent-Hunyuan/HY-Motion-1.0.git
cd HY-Motion-1.0/
git lfs pull
pip install -r requirements.txt
# Inference (Standard / Lite)
python3 local_infer.py --model_path ckpts/tencent/HY-Motion-1.0
python3 local_infer.py --model_path ckpts/tencent/HY-Motion-1.0-Lite
Local execution is suitable for batch generation (many prompts), confidential projects, or when you want to manage outputs yourself. If you only want to “touch it first,” a Hugging Face Space (demo) is also presented, making it easy to try before committing to a local setup.
Prompt Strategy: Short, Specific, and Don’t Try to Get Everything in One Shot
Because HY-Motion 1.0 is text-to-motion, your prompt is the blueprint. But rather than throwing a long “director’s script” at it, it’s often more practical to generate multiple short motions and stitch them—this aligns better with VRAM constraints and iteration. Hugging Face also suggests keeping prompts under 30 words, reinforcing that “short prompts” are pragmatic.
Here are three prompt “templates” that work well in production. These are not model-specific syntax; they’re general instruction-writing patterns you can apply in Japanese or English (and in practice, splitting them into multiple generations is recommended).
Template A: Action + tempo + posture (minimal set)
- Example: “Slowly stand up, face forward, and lightly wave the right hand.”
Template B: Start state → transition → end state (for natural connecting motion)
- Example: “From sitting on a chair, stand up, walk forward two steps, and stop.”
Template C: Add one “intent/emotion” word (to steer performance)
- Example: “Tired: shoulders lowered, heavy steps while walking a short distance.”
Also note where prompts can become too vague. Domestic reporting describes current limitations such as “humanoid only,” and that complex emotional acting, visual attributes like clothing, camera angles, and multi-person interaction are not supported. So specifying detailed cinematography may not translate into the generated skeleton motion.
Ready-to-Use Prompt Examples (Copy/Paste OK), by Production Goal
Below are practical samples organized by goal. The idea is to prepare multiple short instructions rather than betting everything on one long prompt, because that tends to be easier to operate and later stitch together.
1) Game movement + idle (base loop materials)
- “Walk naturally forward three steps, stop, and idle.”
- “Run lightly forward for 2 seconds, slow down, and stop.”
- “Idle: only breathing and weight shift (subtle).”
2) Action (finding a good “hit” for attack/evade)
- “Draw a sword with the right hand, step in half a step, and slash sideways.”
- “Crouch to dodge, then stand up quickly.”
- “Step left once and reset stance.”
Weapon visuals and hitboxes are separate steps, so it’s best to collect patterns where the body flow doesn’t break first. Domestic reporting also gives sword/shield examples.
3) Daily-life gestures (helps immersion for film/VTuber/VR)
- “Sweep the floor: lean forward slightly and move arms slowly.”
- “Read a book: lie on a bed and occasionally turn pages.”
- “Conversation gesture: explain with both hands (not exaggerated).”
These “mundane gestures” are often too small to justify hand-animation from scratch, yet you want them frequently—so generating them as assets can be useful.
Know the Limits First: The Boundary Between What It Can / Cannot Do
HY-Motion 1.0 is not universal. In practice, adoption is easier when you understand the boundaries early and set expectations correctly.
At least the following limitations are explicitly cited in domestic reporting.
- Targets are limited to humanoid characters
- Animals and non-humanoid creatures are not supported
- Complex emotional expression, visual attributes like clothing, camera angles, multi-person interaction, etc. are not supported
Operationally, this suggests the model is strong at generating skeleton motion, but not at specifying “film direction” end-to-end. That’s why it fits best when you rapidly generate motion assets, then refine them later with editing, direction, and implementation.
The Most Important Part: License (Commercial Use / Redistribution / Territory Restrictions), Explained Clearly
HY-Motion 1.0 is openly distributed, but it does not use a general-purpose OSS license like MIT or Apache 2.0. It uses a dedicated community license. On Hugging Face, the license name “tencent-hunyuan-community” is specified, and the LICENSE.txt is publicly available.
Because this area is easy to misunderstand, here are the key practical points (always base final decisions on the actual license text):
1) Territory restriction: EU, UK, and South Korea are excluded
The license explicitly states it does not apply in the EU, UK, and South Korea, defining the Territory as regions excluding them. Usage in those regions may not be justified under the license terms alone, so extra legal/contract caution is needed.
2) Large-scale services require an additional license application (MAU condition)
The license states that if your monthly active users exceed a threshold (the text says “1 million monthly active users” in the prior month at the time of release), you must apply to Tencent for a separate license. Large platforms and popular services integrating the model should treat this as a major risk point.
3) You can’t use it to improve other AI models (including via outputs)
A restriction states you must not use HY-Motion 1.0 results (including outputs) to improve other AI models (with an exception implying HY-Motion itself or derivatives are treated differently). If your plan involves research/data augmentation to train other models, you should check this at design time.
4) Redistribution requires including the license/notice and documenting changes
Conditions for distribution include providing the license text, stating modifications, and including required wording in a Notice file. If your project involves shipping/delivering to third parties, coordinate not just production but distribution operations too.
5) Output rights: Tencent claims no rights to outputs (but responsibility is yours)
The license says Tencent does not claim rights to generated outputs. At the same time, it clarifies that responsibility for using outputs belongs to the user. If you use outputs in production, you still need your standard checks for materials/rights/safety, just like traditional pipelines.
Practical Adoption Plan: A “Small Introduction” That’s Hard to Fail
Rather than putting HY-Motion 1.0 at the center of production immediately, it’s safer to start with a replaceable sub-step. A recommended sequence:
- Start with short motions to understand generation tendencies (walk/stop/sit, etc.)
- Generate a small amount of project-specific gestures (draw weapon, greeting, work motions, etc.)
- Don’t adopt outputs as-is—measure “usable ratio” assuming edits (what percentage becomes usable assets)
- Then scale up (batch inference, templated prompts)
Following VRAM-reduction tips (short prompts, short motion length, seed settings) helps keep environment setup from getting messy.
Summary: HY-Motion 1.0 Is a Practical Model That Shortens the “Entrance” to Motion Creation
HY-Motion 1.0 is released as a family of models that generate skeleton-based 3D human motion from text. Primary sources describe DiT + Flow Matching, scaling to the billion-parameter class, and a staged training pipeline (large-scale pretraining → high-quality fine-tuning → RL with human feedback) to improve instruction-following.
On the other hand, the VRAM requirement is not small, so operational design—“try small, short motions first”—is critical for adoption.
Above all, understanding license constraints (territory restrictions, MAU conditions, restrictions on using outputs to improve other models, etc.) before use is the shortest route to avoiding real production accidents.
Rather than “automating animation,” it’s more realistic to view HY-Motion 1.0 as a tool that increases iteration count in animation production. Try it first on the lightest slice of your most painful motion work.