Download the models
The model weights (~9.3 GB) are not included in the plugin: they are downloaded once, when you ask for it, and shared by all your projects.
Open the panel
Tools > Text To Anim. The panel shows Models needed.
Download
Click Download Models: the editor downloads about 9.3 GB directly from the official Hugging Face
repositories (see Manual download) into
%LOCALAPPDATA%\TextToAnim\Models (or a folder of your choice: Change Folder…). Every file is verified
(SHA-256); the download can be cancelled and resumed.
By downloading you accept the NVIDIA Open Model License and the Meta Llama 3 Community License, linked in the panel.
Ready
When the download ends the model starts by itself. The status turns green (Ready) after 15–30 seconds: you can generate.
Manual download
The models come from the public Hugging Face repositories of the GGUF conversions published by the authors of kimodo.cpp (no account needed). To install them by hand — for example on an offline machine, or if your network blocks the download from the editor — download these three files from the official repositories:
| File | Size | Repository |
|---|---|---|
kimodo-soma-rp-v1.1-f32.gguf (in the models folder of the repository) | 1.1 GB | LocalAI-io/Kimodo-SOMA-RP-v1.1-GGML |
Llama-3-Kimodo-Q8_0.gguf | 8.1 GB | LocalAI-io/Llama-3-Kimodo-GGML |
tokenizer.gguf | 7.3 MB | LocalAI-io/Llama-3-Kimodo-GGML |
Place them in one folder with this layout:
Models/
Llama-3-Kimodo-Q8_0.gguf
tokenizer.gguf
models/
kimodo-soma-rp-v1.1-f32.gguf
then point Change Folder… (or the Models Directory setting) at Models.
The SHA-256 of every file is listed in the MANIFEST.json of each repository, if you want to check them.
Downloading a folder with the editor on a connected machine and copying it works too.
About the background process
The model never runs inside the editor: a crash of the GPU driver or a full graphics memory must not close
your editor and lose your work. The plugin starts a hidden, separate Unreal process
(UnrealEditor-Cmd.exe … -run=TextToAnimBackend) that loads the model and talks to the editor through a
private local pipe — no network port is opened. It closes together with the editor.
You will see it in the Task Manager while the model is loaded. Start Model / Stop Model in the panel load and unload it; loading takes about 10–15 seconds.