P-CORE: Self-Supervised Surface Consistency for Point-Based Neural Editing

Advances in neural rendering enable high-fidelity multi-view reconstruction of 3D scenes, yet free-form non-rigid shape editing remains challenging. Point-based neural representations are attractive for reconstruction because they lack fixed connectivity, so the learned surface topology is not constrained by the initialization. However, this same property causes them to develop holes and surface discontinuities under large deformations. To address this, we propose a self-supervised method that lets point-based representations adapt to large deformations without ground-truth multi-view images of the deformed geometry. The key idea is to generate random deformations and enforce consistency of the predicted surface before and after deformation: the surface predicted from the deformed point cloud should match the deformation applied to the surface predicted from the original point cloud. We build on attention-based point representations, which—unlike splatting-based ones—interpolate between points with a learned kernel rather than a Gaussian around each point; this learned kernel adapts to large deformations without adding or removing points. Experiments on synthetic editing benchmarks (Neural Editor, Objaverse) show that our approach outperforms existing point-based methods in zero-shot editing and substantially reduces artifacts, and qualitative results on DTU and Mip-NeRF 360 demonstrate effectiveness on real-world scenes.