3DGS Generation and Color Editing with Rich Text

3D Gaussian splatting(3DGS) has emerged as a dominant representation for generative 3D modeling, due to its rendering efficiency and quality. However, existing semantic editing methods for 3DGS still face challenges such as low efficiency and poor consistency. 2D diffusion-based editing methods suffer from 3D inconsistency and color bleeding. To address these issues, we propose a rich-text-guided semantic editing approach for 3DGS focusing on color editing. First, we employ an existing text-to-3DGS model to produce initial Gaussians and cross-view attention maps. Second, the attention maps are used to extract semantically consistent 3D segmentations, which establish correspondence between rich-text spans and target editing regions. Finally, we design a region-based diffusion denoising process that incorporates semantic segmentation to blend noise and progressively guide the color of the target regions toward the desired values. Experiments conducted across diverse prompts demonstrate that, compared with two baseline methods, our approach produces more precise editing regions, reduces color bleeding, and yields more natural results. Furthermore, compared with optimization-based method, our method reduces edit time from 10 minutes to 25 seconds per operation.