Seed3D: Semantic-Aware 3D Scene Editing with Diffusion Priors

The ability to edit 3D scenes with precision and realism is essential for applications in autonomous driving, virtual reality (VR), and embodied AI. Traditional scene editing techniques primarily operate at an object level, limiting the scope of fine-grained scene modifications. In this work, we introduce a novel semantic-aware 3D scene editing framework that enables precise and context-sensitive scene alterations while ensuring multi-view consistency. By leveraging implicit scene representations such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3D GS), our approach modifies a user-defined region in a single-view RGB image guided by semantic priors and propagates these modifications across multiple viewpoints to maintain coherence in the 3D space. Furthermore, we explore the application of our method in embodied AI by integrating it with robotic simulators to dynamically refine training environments, enhancing diversity and realism for embodied agents. Experimental results on KITTI-360 and Replica datasets demonstrate that our method outperforms baseline approaches in both content accuracy and cross-view consistency, making it a valuable tool for simulation-driven AI training and real-world deployment.