Researchers Develop Precise Local Editing Method for 3D Gaussian Scenes

Computer vision researchers have introduced a technique that substantially improves the ability to edit individual parts of 3D objects represented as Gaussian Splatting, a modern approach to rendering and manipulating 3D scenes. The work, called RoMaP, tackles two persistent obstacles that have limited practical local editing capabilities.

The Core Challenge#

Editing specific regions of a 3D Gaussian scene requires knowing exactly which pixels belong to which parts across different camera viewpoints. In practice, 2D segmentation methods often produce inconsistent labels when looking at the same object from different angles, making it hard to define clean 3D regions. Additionally, the optimization process used to refine 3D content—known as Score Distillation Sampling (SDS)—is inherently unpredictable, sometimes causing unwanted changes outside the target area or preserving old geometry when it should be modified.

How RoMaP Works#

The framework introduces two key innovations. First, a 3D-Geometry Aware Label Prediction (3D-GALP) module creates robust 3D masks by using spherical harmonics coefficients to model how part labels should vary smoothly across viewpoints. This approach treats part boundaries as having soft transitions rather than hard edges, resulting in consistent segmentations from every angle.

Second, RoMaP modifies the SDS loss function with multiple regularizers. A Scheduled Latent Mixing and Part (SLaMP) editing method generates high-quality edited 2D images and anchors changes tightly to the target region. Additional constraints—such as selectively removing prior information about the original geometry—grant more flexibility to transform parts while preventing stray edits elsewhere.

Impact and Availability#

Experimental results demonstrate state-of-the-art performance on both reconstructed scenes and generated objects, supporting both conservative edits (that preserve surrounding context) and more drastic transformations. The researchers have made code available to support reproducibility and further development.