**Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splatting Editing** **Authors:** Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, Song-Hai Zhang **Institution:** Tsinghua University, Tencent AI Lab, Cardiff University **Abstract:** 3D Gaussian splatting (3D-GS) has gained attention for its capabilities in high-fidelity reconstruction and real-time rendering. However, it couples the appearance and geometry within the Gaussian attributes, limiting the flexibility of editing operations like texture swapping. To address this, the authors propose Texture-GS, which disentangles the appearance from the geometry by representing it as a 2D texture mapped onto the 3D surface. This facilitates appearance editing. The key technical challenge is achieving efficient and smooth texture mapping. The proposed texture mapping module includes a UV mapping MLP to learn UV coordinates for 3D Gaussian centers, a local Taylor expansion for efficient ray-Gaussian intersection mapping, and a learnable texture to capture fine-grained appearance. Extensive experiments on the DTU dataset demonstrate that Texture-GS enables high-fidelity appearance editing and real-time rendering on consumer-level devices, achieving an average rendering speed of 58 FPS on a single RTX 2080 Ti GPU. **Keywords:** Neural rendering, Scene editing, Novel view synthesis, Gaussian splatting, Texture mapping, Disentanglement **Introduction:** The paper addresses the fundamental problem of reconstructing, editing, and rendering photo-realistic scenes, which are crucial in various applications such as film production, computer games, and virtual/augmented reality. Traditional rendering pipelines use polygonal meshes, which are efficient for rendering and editing. However, 3D Gaussian Splatting (3D-GS) offers a more flexible representation for complex scenes, but it couples appearance and geometry, hindering editing operations. Texture-GS aims to disentangle these components, enabling various editing applications. **Method:** Texture-GS uses a UV mapping MLP to project 3D points into 2D UV space and a learnable 2D texture to represent the appearance. The UV mapping MLP is learned using cycle-consistency constraints and a Taylor expansion for efficiency. The texture value is learned using a photometric loss and the Taylor expansion-based approximation of the MLP. **Experiments:** Experiments on the DTU dataset show that Texture-GS recovers high-quality 2D textures and enables various editing applications, including global texture swapping and fine-grained texture editing. The method achieves real-time rendering speed on consumer-level devices. **Contributions:** - First to disentangle geometry and texture in 3D-GS. - Proposes a novel texture mapping module for efficient and smooth texture mapping. - Validates the effectiveness of Texture-GS for novel view synthesis, texture swapping, and local appearance editing. **Related Work:**Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splatting Editing** **Authors:** Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, Song-Hai Zhang **Institution:** Tsinghua University, Tencent AI Lab, Cardiff University **Abstract:** 3D Gaussian splatting (3D-GS) has gained attention for its capabilities in high-fidelity reconstruction and real-time rendering. However, it couples the appearance and geometry within the Gaussian attributes, limiting the flexibility of editing operations like texture swapping. To address this, the authors propose Texture-GS, which disentangles the appearance from the geometry by representing it as a 2D texture mapped onto the 3D surface. This facilitates appearance editing. The key technical challenge is achieving efficient and smooth texture mapping. The proposed texture mapping module includes a UV mapping MLP to learn UV coordinates for 3D Gaussian centers, a local Taylor expansion for efficient ray-Gaussian intersection mapping, and a learnable texture to capture fine-grained appearance. Extensive experiments on the DTU dataset demonstrate that Texture-GS enables high-fidelity appearance editing and real-time rendering on consumer-level devices, achieving an average rendering speed of 58 FPS on a single RTX 2080 Ti GPU. **Keywords:** Neural rendering, Scene editing, Novel view synthesis, Gaussian splatting, Texture mapping, Disentanglement **Introduction:** The paper addresses the fundamental problem of reconstructing, editing, and rendering photo-realistic scenes, which are crucial in various applications such as film production, computer games, and virtual/augmented reality. Traditional rendering pipelines use polygonal meshes, which are efficient for rendering and editing. However, 3D Gaussian Splatting (3D-GS) offers a more flexible representation for complex scenes, but it couples appearance and geometry, hindering editing operations. Texture-GS aims to disentangle these components, enabling various editing applications. **Method:** Texture-GS uses a UV mapping MLP to project 3D points into 2D UV space and a learnable 2D texture to represent the appearance. The UV mapping MLP is learned using cycle-consistency constraints and a Taylor expansion for efficiency. The texture value is learned using a photometric loss and the Taylor expansion-based approximation of the MLP. **Experiments:** Experiments on the DTU dataset show that Texture-GS recovers high-quality 2D textures and enables various editing applications, including global texture swapping and fine-grained texture editing. The method achieves real-time rendering speed on consumer-level devices. **Contributions:** - First to disentangle geometry and texture in 3D-GS. - Proposes a novel texture mapping module for efficient and smooth texture mapping. - Validates the effectiveness of Texture-GS for novel view synthesis, texture swapping, and local appearance editing. **Related Work: