Title: GSDeformer: A Real-Time, Direct, and Extensible Framework for Cage-Based Deformation in 3D Gaussian Splatting

Abstract

This paper introduces GSDeformer, a novel technique that facilitates cage-based deformation within the 3D Gaussian Splatting (3DGS) framework. By employing a proxy point-cloud representation, our method effectively links cage-based deformation with 3DGS. Specifically, this point cloud is derived from 3D Gaussians, allowing any manipulations applied to the cloud to be converted into corresponding transformations for the underlying Gaussians. To address potential distortion or bending resulting from these deformations, we integrate a splitting mechanism that serves as an effective approximation.

Crucially, GSDeformer operates without altering or expanding the fundamental architecture of 3D Gaussian Splatting. This design choice ensures seamless compatibility with any pre-trained vanilla 3DGS model or its derivatives. Furthermore, we propose an automated cage construction process for 3DGS and related variants, leveraging a render-and-reconstruct strategy. Experimental evaluations confirm that GSDeformer outperforms current methods in terms of deformation quality and exhibits robustness even under extreme manipulations. The system requires no retraining for editing tasks, supports real-time performance, and offers extensibility to other 3DGS variants.

Project Page: https://jhuangbu.github.io/gsdeformer/

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC