You Can Now Generate Animations for FREE in Unreal Engine 5.9!

Summary

The video demonstrates an experimental AI text-to-animation workflow in Unreal Engine 5.9 Preview 1. Using NVIDIA’s animation-generation plugins, developers can describe a character action in text and generate an animation locally, without paying per-generation credits or leaving the editor.

Setup and tutorial steps

1. Install Unreal Engine 5.9 Preview 1 and create a project, such as the third-person template.
2. Enable the NVIDIA Omniverse and Anim Gen Offline plugins, then restart the editor.
3. Create a free Hugging Face account, request access to the Meta Llama 3 Instruct model, create a read-access token, and enter it in the Omniverse settings.
4. Check that the required Python packages, including PyTorch and other dependencies, are installed. The presenter notes that missing modules can cause errors and provides dependency instructions in the video description.
5. Create a Level Sequence, add the default Soma character, and optionally select a project character’s skeletal mesh for retargeting.
6. Enter a text prompt, set the animation’s start and end range, choose a quality level, and generate the animation. The quality settings correspond to 20, 50, or 100 diffusion steps; higher quality takes longer.
7. After generation, play the animation and adjust its movement path using the editable spline, or delete the spline if it is not needed.
8. Right-click the character and select Bake Animation Sequence to export the result as an animation asset.

The first generation requires downloading a model file of about 17 GB. The presenter says NVIDIA’s standalone implementation can use under 3 GB of video memory, with the text encoder running on the CPU. They also state that the models are commercially licensed for use in released projects.

Results and assessment

Examples include a slow, arms-out zombie walk, a character waving, and jumping jacks. Prompts can be revised—for example, by specifying which hand should wave—and the animation can then be regenerated.

Results vary: some actions look better than others, and not every animation is convincing. The feature is presented primarily as a way to prototype, create placeholder animations, and test ideas quickly, rather than as a guaranteed replacement for polished animation or motion capture.

Main speaker/source: The Smart Poly presenter, demonstrating the Unreal Engine plugins and describing NVIDIA’s animation tools, Hugging Face access, and Meta Llama model.