RoMaP: Robust 3D-Masked Part-level Editing in 3D Gaussian Splatting with Regularized Score Distillation Sampling

Hayeon Kim1, *, Ji Ha Jang1, *, Se Young Chun1, 2
1Dept. of Electrical and Computer Engineering, 2INMC & IPAI *These authors contributed equally to this work
Seoul National University, Korea
ICCV 2025
arXiv Code

Video

Motivation

Limitations of prior local 3D editing methods leveraging 2D part-level segmentation and edits. Our approach enables accurate and view-consistent 3D part editing, overcoming the limitations of 2D segmentation-based methods..

Abstract

Due to limited 3D data, recent prior arts in Recent advances in 3D neural representations and instance-level editing models have enabled the efficient creation of high-quality 3D content. However, achieving precise local 3D edits remains challenging, especially for Gaussian Splatting, due to inconsistent multi-view 2D part segmentations and inherently ambiguous nature of Score Distillation Sampling (SDS) loss. To address these limitations, we propose RoMaP, a novel local 3D Gaussian editing framework that enables precise and drastic part-level modifications. First, we introduce a robust 3D mask generation module with our 3D-Geometry Aware Label Prediction (3D-GALP), which uses spherical harmonics (SH) coefficients to model view-dependent label variations and soft-label property, yielding accurate and consistent part segmentations across viewpoints. Second, we propose a regularized SDS loss that combines the standard SDS loss with additional regularizers. In particular, an L1 anchor loss is introduced via our Scheduled Latent Mixing and Part (SLaMP) editing method, which generates high-quality part-edited 2D images and confines modifications only to the target region while preserving contextual coherence. Additional regularizers, such as Gaussian prior removal, further improve flexibility by allowing changes beyond the existing context, and robust 3D masking prevents unintended edits. Experimental results demonstrate that our RoMaP achieves state-of-the-art local 3D editing on both reconstructed and generated Gaussian scenes and objects qualitatively and quantitatively, making it possible for more robust and flexible part-level 3D Gaussian editing.

Method

Overall pipeline of RoMaP. RoMaP first segments 3D Gaussian using 3D-GALP, leveraging the soft-label properties of Gaussians to address the intricacies of part-level segmentation. With anchors consisting of both label-consistent and inconsistent Gaussians, we refine 3D segmentation considering locality with neighboring Gaussians. Then, in local 3D editing, we remove Gaussian priors and introduce a new direction of modification using SLaMP-edited images, followed by refinement through a close examination of Gaussians.

Segmentation Qualitative Result

3D Gaussian segmentation results of 3D-GALP With our 3D-GALP, 3D Gaussian segmentation accurately captures diverse object parts, addressing the limitations of 2D part segmentation and the inherent mixed nature of 3D Gaussian segmentation labels.

Editing Qualitative Result

Enhanced controllability in 3D local edits with RoMaP Our approach enables precise manipulation of specific 3D parts. As shown above, RoMaP provides diverse control over multiple narrow regions within a single 3D object, allowing deformations in targeted areas like a ‘duck’s beak’ or ‘jellyfish hair’ and facilitating various modifications in targeted area such as a lamp’s lampshade.

BibTeX

@inproceedings{kim2025romap,
    title={Robust 3D-Masked Part-level Editing in 3D Gaussian Splatting with Regularized Score Distillation Sampling},
    author={Hayeon Kim, Ji Ha Jang, Se Young Chun},
    booktitle={International Conference on Computer Vision (ICCV)},
    year={2025}
  }