Proxy-Free Gaussian Splats Deformation with Splat-Based Surface Estimation

TL;DR

This paper introduces SpLap, a proxy-free deformation framework for Gaussian splats that relies on a surface-aware splat graph to define a robust Laplacian on the splat means. By constructing connectivity from splat intersections and using geodesic distances, SpLap performs Laplacian-based edits via ARAP or BBW without external proxies, followed by a surface-preserving kernel adaptation to maintain accurate surface coverage. Empirically, SpLap achieves deformation quality near ground truth on a 50-object benchmark spanning diverse categories and also generalizes to other surface-aligned GS methods, outperforming both proxy-based and proxy-free baselines in 3DPCK and visual fidelity. This approach broadens the utility of Gaussian splatting, enabling high-fidelity, large-scale edits directly in the splat domain with reduced reliance on costly surface reconstructions. The framework's reliance on the intrinsic surface structure of GS makes it a practical, scalable solution for editing 3D scenes captured from 2D images.

Abstract

We introduce SpLap, a proxy-free deformation method for Gaussian splats (GS) based on a Laplacian operator computed from our novel surface-aware splat graph. Existing approaches to GS deformation typically rely on deformation proxies such as cages or meshes, but they suffer from dependency on proxy quality and additional computational overhead. An alternative is to directly apply Laplacian-based deformation techniques by treating splats as point clouds. However, this often fail to properly capture surface information due to lack of explicit structure. To address this, we propose a novel method that constructs a surface-aware splat graph, enabling the Laplacian operator derived from it to support more plausible deformations that preserve details and topology. Our key idea is to leverage the spatial arrangement encoded in splats, defining neighboring splats not merely by the distance between their centers, but by their intersections. Furthermore, we introduce a Gaussian kernel adaptation technique that preserves surface structure under deformation, thereby improving rendering quality after deformation. In our experiments, we demonstrate the superior performance of our method compared to both proxy-based and proxy-free baselines, evaluated on 50 challenging objects from the ShapeNet, Objaverse, and Sketchfab datasets, as well as the NeRF-Synthetic dataset. Code is available at https://github.com/kjae0/SpLap.

Figures (10)

• Figure 1: SpLap is proxy-free Laplacian-based deformation framework for Gaussian splats. Our method enables high-quality, large-scale deformation, all without relying on external proxy. We show results for both As-Rigid-As-Possible ARAP (top row) and Bounded Biharmonic Weights BBW (bottom row) deformations. The red dots indicate the interaction handles and arrows show the direction of the applied edit.
• Figure 2: Illustration of different connectivity search methods. Existing metrics, (a) and (b), often produce incorrect connections (red) or fragmentation (gray) due to the lack of structural prior. In contrast, by leveraging geometric constraints of surface-aligned GS, our method avoids this suboptimal situations like (c).
• Figure 3: Schematic of kernel adaptation methods. Simply estimating local transformation fails due to a scale mismatch, leading to visual artifacts. Our surface-preserving method avoids this by directly maintaining the original surface coverage.
• Figure 4: Qualitative comparison of deformation results. The red dots denote the interaction handles, with arrows indicating the edit direction. The top five rows show results using the ARAP ARAP, while the bottom three rows use BBW BBW. Our method demonstrates superior visual fidelity and geometric consistency compared to baselines, especially in preserving fine details and preventing artifacts across both deformation techniques.
• Figure 5: Comparison of neighborhood estimations. Graph distance based method provides improved robustness than raw graph connectivity on challenging geometries, such as thin structures.