3DTexMOR: 3D Gaussian Multi-object Removal via Texture-Space Inpainting

3D object removal aims to remove target objects from reconstructed scenes while completing the geometry and appearance of the regions they occlude. Existing methods built on Neural Radiance Fields (NeRF) or 3D Gaussian Splatting (3DGS) typically inpaint 2D images and use the completed views to guide 3D scene completion. While these methods achieve promising results in scenes with few target objects, they often struggle with complex multi-object layouts, where individual views capture only part of the surrounding texture and structure needed to guide inpainting. These limited visual cues make 2D inpainting prone to artifacts. Moreover, appearance inconsistencies across inpainted views can introduce conflicting supervision for 3D scene completion, leading to blurry reconstructions. To address these challenges, we propose 3D Gaussian Multi-Object Removal via Texture-Space Inpainting (3DTexMOR). Our key idea is to perform inpainting in a unified texture space that combines complementary observations across views, providing richer visual information for recovering missing regions while promoting cross-view appearance consistency. Specifically, we introduce a texture-space inpainting module that aggregates multi-view observations into texture maps and reprojects the completed maps into camera views to supervise 3D Gaussian scene completion. Since direct RGB aggregation is unreliable on glossy surfaces, we further decompose appearance and aggregate view-independent intrinsic attributes in texture space. Finally, because completed textures alone do not sufficiently constrain 3D structure, we introduce a geometrically regularized Gaussian completion module to encourage plausible geometry in the completed regions. Extensive experiments demonstrate that 3DTexMOR produces visually plausible scene completions and achieves state-of-the-art performance on multi-object removal, improving PSNR by 5.8 dB and reducing LPIPS by at least 22% compared with existing methods.