Hierarchical Diffusion-Based Text-to-Humanoid Motion Generation Technology

Text-to-Humanoid Motion Generation Technology Combining Hierarchical Diffusion Models and Physics-Based Correction

This technology automatically generates humanoid motion from text by extracting and decomposing verbs into behavioral units, generating motion via conditional diffusion models, stitching segments using overlapping interpolation, and integrating physics-based correction into the reverse diffusion process.

Existing text-to-humanoid motion generation methods often produce unnatural results, struggling with seamless transitions between movements and failing to account for physical constraints such as joint range of motion and foot contact.

By utilizing behavioral decomposition, conditional diffusion models, overlapping interpolation, and integrated physics correction, this technology enables the automatic generation of realistic motion from text alone, making it ideal for 3D character and humanoid content production.

‍

‍

Key Features:

• A user terminal that accepts text input and outputs a 3D humanoid object performing the motion described in the text.
• A behavioral decomposition unit that extracts verbs from the text, breaks them down into behavioral units, and infers the sequence and timing of the actions.
• A motion generation unit that calls a pre-built conditional diffusion model to generate motion corresponding to each behavioral unit.
• A stitching unit that joins motion segments using overlapping interpolation, and a physics-based correction unit that integrates constraints such as joint range of motion, velocity, and foot contact into the reverse diffusion process.

‍

This patent portfolio consists of four family patents covering AI-based 3D motion generation technology developed by Bumblebee Co., Ltd. Technology transfer is available alongside the patents, with pricing negotiable. Please refer to the attached technical manual for further details.

‍