Abstract
Abstract In recent years, the latest advances in the field of 3D content creation have been primarily based on neural radiation fields (NeRF). Despite significant research achievements, these methods often suffer from slow optimization speeds and high memory consumption. More importantly, these studies still mainly focus on static modeling, with very limited research on dynamic models, which restricts their practical applications. This paper proposes a method Yibaogaussian for generating dynamic 3D models from text descriptions, which can generate dynamic 3D models while ensuring efficiency and quality. Our key insight is on constructing canonical space and temporal deformation field . We build static 3D Gaussian fields in canonical space to capture the overall geometry and appearance features of a scene. By dynamically mapping the parameters of the Gaussian field to the temporal deformation field, we achieve continuous changes and natural motion of the static 3D model over time. Extensive experiments demonstrate that our proposed method offers higher efficiency and highly competitive generated quality. Notably, YibaoGaussian can generate high-quality dynamic 3D models from text descriptions in just 3 minutes, approximately 3 times faster than existing methods. This study verifies the feasibility of combining Gaussian representation and temporal deformation field in regular space, providing a new solution for text-driven dynamic 3D model generation and laying the technical foundation for multimodal applications such as digital humans, virtual reality (VR/AR), and film and television special effects.