LDS3D: Language‐Driven Object Stylization based on 3DGS

Abstract 3D Gaussian Splatting (3DGS) has emerged as a powerful method that offers significant improvements in both rendering quality and efficiency. However, the task of achieving open‐ended scene style transfer remains a considerable challenge, as existing methods typically focus on the entire scene and fail to address the need for fine‐grained editing of individual objects within the 3D scene. To address this, we introduce LDS3D, a novel language‐driven method for stylization of 3D scene objects. This approach enables the open‐ended stylization of specific objects within a scene by natural language prompts. First, we present a language‐driven segmentation method, which accurately integrates multi‐view consistent CLIP latent features of scene objects into 3DGS. Subsequently, we train a decoder to map these latent features back into high‐dimensional CLIP features, enabling open‐ended vocabulary segmentation of the scene. Following this, we propose an object stylization strategy that leverages both CLIP loss and content loss, employing open‐ended language prompts to fine‐tune the spherical harmonic coefficients of the selected Gaussians, thus achieving style transformation that aligns with the language requirements. LDS3D allows for the identification and stylization of designated objects in a scene by natural language, enabling precise and flexible style transfer. Moreover, our method exhibits strong responsiveness to complex semantic prompts, capturing and effectively applying intricate style attributes to target objects. We demonstrate the effectiveness of our method across a diverse range of 3D scenes and complex language prompts, showcasing its capacity for flexible and accurate object‐level stylization in 3DGS.