Type a prompt, get human motion on your character.
This is a Unity front end for NVIDIA Kimodo, a
text-to-motion diffusion model: generate from a prompt, preview it live on any Humanoid rig, shape it
with Scene-view constraints, sequence a whole shot on a timeline, and bake the result to an
ordinary AnimationClip that needs none of this at runtime.
Independent wrapper by AminHP — not affiliated with, sponsored by, or endorsed by NVIDIA. Apache-2.0. See THIRD-PARTY-NOTICES.
Every feature below, in a couple of minutes — starting the server, a prompt, constraints, the timeline, freezing, and baking.
Window ▸ Package Manager ▸ + ▸ Install package from git URL:
https://github.com/Amin-HP/Kimodo-Unity.git?path=/Packages/com.aminhp.kimodobridge
Append a tag to pin a version, e.g. …com.aminhp.kimodobridge#v0.2.8. Only the package folder is
fetched — the Unity project in this repository is just where it is developed.
You also need Python 3.10 with Kimodo installed and a CUDA GPU (~8 GB VRAM is enough). The bridge server that wraps Kimodo ships inside the package.
- GameObject ▸ Kimodo ▸ Bridge Manager → in Setup, point Python at the environment that has Kimodo → Start. It connects on its own.
- Select a Humanoid character → GameObject ▸ Kimodo ▸ Set Up Selected Character.
- Type a prompt, press Generate, watch the progress bar, scrub the preview.
- Bake to AnimationClip when you like it.
The server can also be run by hand (python -m kimodo_bridge) or on another machine — it is plain HTTP.
→ The full manual is the package README: constraints, the timeline, freezing, baking, the runtime API and troubleshooting.
- Text → motion, retargeted onto any Unity Humanoid through Mecanim muscle space.
- Live preview (play / scrub) and bake to a reusable
.anim, in place or travelling, with the root-motion scale measured automatically per character. - Constraints, authored with Scene-view gizmos and a dockable tool overlay:
- Waypoints — the ground path the pelvis follows, with per-point facing.
- End-Effectors — hand / foot keyframes; drag a limb and it bends to reach.
- Full-Body poses — pose the whole rig, or switch individual joints off.
- Each key is shown as a transparent ghost of your own model, faded down to the part of the body the key is about, so you place a hand on an arm rather than on a bare skeleton.
- Timeline — sequence a shot as ordered prompt segments on one frame axis carrying every constraint key, give a beat its own constraint strength, and ❄ freeze the segments you are happy with so Generate skips them entirely and only regenerates the rest.
- Generation progress read from the model's own denoising loop, not guessed from elapsed time.
- Survives recompiles — the generated motion is cached, so a script change no longer costs you a minute of regenerating.
| Path | What it is |
|---|---|
Packages/com.aminhp.kimodobridge/ |
The package — runtime + editor code, and the Python bridge server under Server~/. This is what the git URL installs. |
Assets/ |
A minimal URP project used to develop and test the package. The test character is not committed. |
- Unity 6000.0+, and a character with Rig ▸ Animation Type ▸ Humanoid.
- Python 3.10 with NVIDIA Kimodo installed (
pip install -e .in its repo). - A CUDA GPU; ~8 GB VRAM works with the text encoder on the CPU (the default).
Apache-2.0 for this repository's own code. NVIDIA Kimodo is separate, under its own Apache-2.0 license, and is not redistributed here — install it yourself.
If this saved you some time, you can support the work here: Buy me a coffee.