I've seen NVIDIA's CALM before, and today I'll introduce a project trained on the basis of MotionGPT specifically for gaming. AnimationGPT is a project focused on generating combat-style character animations from text.
The generated animations can be downloaded in mp4 or bvh formats.
To incorporate large-scale language data and models into motion generation tasks, AnimationGPT adopts a unified motion-language framework called MotionGPT. Specifically, MotionGPT uses a discrete vector quantization method for human movements and converts 3D motions into motion tokens, similar to the process of generating word tokens. Based on this "motion vocabulary," language modeling is performed uniformly for both motion and text, treating human motion as a new foreign language. The model then translates human natural language into the most appropriate sequence of human actions through its inference capabilities.
This project was trained on the basis of MotionGPT and generated the first dataset specifically for combat-style character animations, named CombatMotion. Compared with current text-to-motion datasets, CombatMotion has the following characteristics:
Derived from game assets. Features combat styles, with a concentration on animation styles in action games and a bias towards certain types of actions. More detailed textual annotations.
The dataset can also be downloaded here👇: https://drive.google.com/file/d/148AwoovJrnh4F0q_HbU83WCWwFooLSZj/view