Text to motion AI converts a written description of a human action into animated joint data — a skeleton rotating and moving over time — rather than into pixels. You describe a movement in plain language, the model produces a motion draft, and you export it as FBX or JSON for use in a 3D pipeline. The results are strongest when you treat them as blocking: a fast, specific first pass that still needs retargeting and polish.
This guide covers what text-to-motion actually outputs, the prompt habits that improve results, five concrete prompt examples you can adapt, and the cleanup workflow that turns a draft into usable animation.
What Text to Motion AI Actually Outputs
The output of a text-to-motion model is not an image, a video, or a finished character. It is animation data: a humanoid skeleton whose joints have rotations (and often root translation) keyed across a timeline. Two formats dominate the practical workflow.
FBX is a widely supported interchange format for animation and game engines. Blender, Unity, and Unreal all import it, which makes it the default choice when you are moving motion into a production tool. You can preview and convert it if needed using the 3D viewer and the 3D converter.
JSON is the raw-data option. It is useful when you want to inspect joint values directly, write custom tooling, or feed the animation into a pipeline that expects structured data rather than a binary file.
What neither format contains: your character. The model produces motion applied to a skeleton. Putting that motion on your specific character is retargeting, and it happens in Blender, Unity, or Unreal after you download.
| Aspect | FBX output | JSON output |
|---|---|---|
| Primary use | Importing into Blender, Unity, Unreal | Inspecting or scripting raw joint data |
| Carries a skeleton | Yes, with keyed animation curves | Yes, as structured joint values |
| Human-readable | Not easily | Yes, with a text editor |
| Typical consumer | Animation and game pipelines | Custom tooling and automated pipelines |
| Native engine import | Supported by all three major tools | Requires your own parser |
The practical rule: download FBX unless you have a specific reason to want raw data.
Why motion data and not video
Motion data is composable in a way that video is not. Once you have the animation, you can retarget it to any humanoid, change the camera, re-light the scene, blend it with other clips, and retime it to a beat. A rendered clip gives you the look of one shot; motion data gives you an asset you can reuse. That is why text-to-motion sits naturally inside game and 3D animation pipelines rather than replacing them. For a broader view of where it fits, see text to motion AI use cases.
What Prompts Work
Prompting for motion is different from prompting for images. The model is trying to schedule human movement, so the most useful thing you can do is be specific about the movement and silent about everything else.
One human action per prompt
Each prompt should describe a single, coherent action with a clear beginning and end. "A person jogs three steps and leaps over a gap" is one action. "A person jogs, answers a phone, then sits down" is three actions, and the model may blend them awkwardly or drop one. If you need a sequence, generate separate clips and blend them in your editor.
Keep it under about sixty words
Long prompts do not add fidelity; they add ambiguity. A short, clear description of the movement outperforms a paragraph. Aim for one or two sentences that answer: who moves, how, in what direction, and what happens at the end.
Describe motion, not appearance
Words about clothing, hair color, age, or setting are wasted — the model does not produce appearance. Spend every word on the body: limbs, torso, direction, speed, and the shape of the action. "A person in a red hoodie" tells the model nothing useful; "a person raises both arms overhead and stretches" tells it everything.
Name the direction and the ending
The two most common sources of unusable drafts are ambiguous direction and unclear endings. Say "walks forward," "turns to the left," or "steps backward," and describe how the action resolves — "then comes to a stop," "then returns to standing." A clear ending makes the clip far easier to blend.
Use ordinary, concrete verbs
Walk, run, jump, crouch, reach, turn, sit, throw, climb, push, pull, kick. Concrete verbs map cleanly to motion patterns. Vague phrasing like "does something athletic" or "moves dynamically" gives the model nothing to schedule.
Five Concrete Prompt Examples
Adapt these directly or use them as templates. Each describes one action, stays short, and focuses on movement.
-
"A person walks forward at a steady pace, stops, and turns their head to look over their right shoulder, then holds the pose." One action with a clear resolution — useful for idle-to-look transitions.
-
"A person crouches down, then springs upward into a vertical jump, landing on both feet with knees bent." Tests the model's crouch-to-jump timing and landing compression.
-
"A person raises both arms overhead, leans back slightly, and stretches, then lowers their arms to their sides." A simple, symmetric action that is easy to evaluate for limb intersections.
-
"A person jogs forward, plants their left foot, and turns ninety degrees to the right without stopping." Combines locomotion with a directional change, which is a common requirement in games.
-
"A person reaches forward with their right hand, grabs something at chest height, and pulls it back toward their chest." A non-locomotive interaction that exercises upper-body motion.
Notice that none of these mention appearance, mood, or setting. That is deliberate. Once these baseline prompts produce clean drafts, you can add complexity — speed, direction changes, or a follow-up beat — one variable at a time.
The Cleanup Workflow
Every draft needs finishing. The order below is the one that wastes the least time, because each stage depends on the previous one being settled.
Step 1: Review at speed, then from several angles
Watch the clip at normal playback speed first; timing problems are invisible in still frames. Then orbit it in 3D to catch limb intersections, a twisted spine, or a root that drifts. A 3D viewer is enough for this inspection pass.
Step 2: Import and retarget
Bring the FBX into Blender, Unity, or Unreal and map its bones to your character's skeleton. Expect differences in rest pose, proportions, and orientation. Retargeting is where most of the practical effort lives; the FBX animation generator page goes deeper on engine specifics.
Step 3: Fix ground contact and root motion
Generated motion often slides relative to the floor or drifts in space. Correct foot contact, and decide whether the clip should drive the character (root motion) or be driven by gameplay code. This is the step that most often makes the difference between "looks like a glitch" and "looks intentional."
Step 4: Retime and blend
Generated clips are discrete actions, while performances are continuous. Retime sections to hit your beats, and blend transitions between clips so the character does not pop between actions.
Step 5: Polish
Hand adjustments to hands, spine, weight shifts, and secondary motion are what make animation feel alive. This is skilled animator work, and there is no shortcut — but starting from a generated draft means the polish pass begins from a coherent action instead of a blank timeline.
Why Start From a Draft Rather Than Blank
The value of text-to-motion is not that it replaces the animator; it is that it replaces the blank timeline. Blocking out a walk, a jump, or a turn by hand takes time before you even know whether the action reads. A generated draft gives you a concrete version to react to in minutes, which changes the conversation from "what should this look like?" to "what should we adjust?" For a full browser workflow, see how hunyuanmotion.org generates motion; for the underlying capability, the text3d page is a starting point. Credits are available as one-time packs, and signed-in accounts can claim free credits daily — see pricing.
hunyuanmotion.org is an independent interface and is not affiliated with Tencent.
FAQ
What is text to motion AI in one sentence?
It is a system that takes a written description of a human action and produces animated joint data — a moving skeleton — which you can export as FBX or JSON and apply to a 3D character. It produces movement, not images, video, or a finished character.
What file format should I download, FBX or JSON?
Choose FBX if you are importing into Blender, Unity, or Unreal, because those tools read it natively and it carries the animation cleanly. Choose JSON if you want to inspect or transform the raw joint values, build custom tooling, or feed a pipeline that expects structured data.
Why does my prompt produce a strange or blended motion?
The most common cause is asking for more than one action in a single prompt. Text-to-motion models schedule one coherent movement well; when a prompt contains several actions, the result can blend or drop parts. Split the sequence into separate clips and blend them afterward, and keep describing motion rather than appearance.
How long should a prompt be?
Under roughly sixty words, and ideally one or two sentences. The goal is to answer four questions: who moves, how they move, in what direction, and how the action ends. Longer prompts tend to add ambiguity rather than fidelity.
Is the generated motion production-ready?
No, and expecting that leads to disappointment. The draft is a strong blocking pass: readable, specific, and fast to produce, but it still needs retargeting to your skeleton, foot-contact and root-motion fixes, blending, retiming, and hand polish.
Do I need a 3D program at all?
To evaluate a draft, no — a browser-based 3D viewer is enough to watch and inspect the motion. To put the motion on your own character and ship it, yes: retargeting and cleanup happen in Blender, Unity, Unreal, or a similar tool.