NVIDIA Kimodo
Motion · local
Runs in Docker on your own GPU. Free, seeded, and fast enough to regenerate a clip from its definition instead of storing the result.
Around fourteen seconds for a fresh prompt, measured on our own card.
Integrations
Looking for the things that bind these to a game framework? Add-ons.
What the plugins talk to.
Providers and runtimes: the models that generate, the compute that runs them, and the protocol an agent reaches all of it through. Choose one per asset rather than one per project.
Why providers are pluggable
Today’s model is the worst one we will ever ship on.
There are two ways to build on top of AI: assume it is as good as it will get and squeeze it, or assume it is the floor and put the effort into the harness. We build the second way, and this page is what that looks like from outside.
No stage hard-depends on a particular model, every stage records what produced its output, and a provider is a registration rather than a fork. An improvement arrives as a config change, not a rewrite.
Providers and runtimes
Uthana
Motion · hosted
Hosted motion generation behind the same provider interface as the local runner. For when you do not own a card, or do not want to run one.
RunPod
Compute · rented
Rent a GPU from inside the editor and shut it down when the batch finishes. One key, shared by every set that rents hardware — entered once, wherever you happen to be.
Motion clips from a rented pod match local output bone for bone; a rented face runner pulls the same container image a desktop builds.
ElevenLabs
Speech · hosted
Bring your own key and your own account. We do not resell speech. The ledger records the model used for every line.
It reports its seed as best-effort, so output is hashable and never reproducible.
Unreal Engine audio-to-face
Face · local
The engine's own audio-driven solver. Runs locally on CPU, with no account and no per-use cost.
No MetaHuman licence is required for the solve itself — roughly one second of solve per four seconds of audio.
Model Context Protocol
The seam
Built natively on UE 5.8's own ModelContextProtocol and toolset registry plugins, rather than as an external tool bridged into Unreal.
Which is why a toolset is a plugin like any other, and why installing one adds tools with no edit to ours.
NVIDIA Audio2Face-3D
Face · self-hosted or rented
Open-sourced models that self-host and need no engine to solve, so the same runner serves an editor, a batch or a browser. Lip-sync needs no account at all — the face models are ungated and ship inside the container image.
Runs on your own NVIDIA card through Docker, or on a GPU rented by the hour from inside the editor. The engine solver stays the free local one.
NVIDIA Audio2Emotion
Face · self-hosted
Reads the emotion out of the delivery, frame by frame, and drives the face with it — so a line performed angry looks angry without anyone labelling it. An emotion you name yourself still works, and the two combine: take it from the recording, then lean it further.
The one gated piece. A free Hugging Face token and one licence click; without it every face still solves, with the emotion you give it.
NVIDIA ACE and open-weight speech
Speech · hosted and local
Behind the same provider interface, exactly as the local motion runner now sits beside the hosted one.
Coming
Two hard boundaries
No tool writes a credential. Signing in is a human action, performed once in project settings.
Agents can ask whether a key exists. They can never set one or read one. No tool can delete a voice on a provider either: a key that can create a voice can usually delete one, and deleting a voice loses every line it ever spoke.