Design and implementation of a diffusion model for text conditioned motion generation
Loading...
Date
2025-10-13
Authors
Φραγκισκάτος, Χαράλαμπος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Motion synthesis poses one of the most challenging goals of computer animation. It is inherently complex and our sensitivity to unrealistic cues makes it a challenge for animators and graphic professionals worldwide. At the same time advancements in technology and current trends have created a high demand for plausible digital human motion, raising the need to automate this task. Natural language is the simplest method to explain motion and for this reason a text-conditioned motion synthesis algorithm seems like a promising option due to its ease of use.
Many approaches have tackled the task of motion synthesis conditioned on text with plenty of them utilizing the diffusion framework owing to its success in image and video generation. Most of them are of limited use due to their long inference times and the data-driven method they follow introduces many artifacts that degrade the quality of the output motion. This thesis explores diffusion as a method of motion synthesis. Our goal is to design and implement a machine learning model able to produce motion from descriptive text inputs at fast generation times that also supports motion guidance through predefined poses.
Description
Keywords
Diffusion, Motion synthesis, Text-conditioned generation, Unet, Generative AI, Machine learning, Keyframe imposing