In this paper, we propose RoMaP, a novel framework for precise local editing of 3D models based on Gaussian Splatting. To address the inaccurate 2D segmentation of conventional Gaussian Splatting and the ambiguity of Score Distillation Sampling (SDS) loss, RoMaP generates precise and consistent segmentations via the 3D-Geometry Aware Label Prediction (3D-GALP) module, and performs precise editing of the target region via the regularized SDS loss function and the Scheduled Latent Mixing and Part (SLaMP) editing method. SLaMP produces high-quality, locally edited 2D images while maintaining contextual consistency, and additional regularization terms (e.g., removing the Gaussian prior) enhance the flexibility by allowing changes beyond the original context. Experimental results demonstrate that RoMaP achieves state-of-the-art local 3D editing performance on both reconstructed and generated Gaussian scenes and objects.
