Geometry-Aware Style Transfer in 3D Gaussian Splatting

Traditional 3D style transfer methods often struggle to balance appearance changes with structural integrity. Most existing approaches for 3D Gaussian Splatting (3DGS) prioritize color-based stylization while keeping geometry fixed, as direct manipulation of geometric parameters often leads to structural collapse or visual artifacts. This paper asks: how can we effectively incorporate geometric transformation into 3D style transfer without sacrificing the stability of the underlying 3D scene representation?

The authors propose a decoupled optimization scheme that treats color and geometry as distinct but interacting components. Instead of updating all parameters simultaneously, the framework alternates between a color optimization phase and a geometry optimization phase. This separation prevents the interference that typically occurs when color-based stylistic patterns are forced onto geometric primitives.

To guide this process, the authors introduce Geometry-Aware Contrastive Feature Matching (GCFM). Unlike standard feature matching that relies solely on color, GCFM integrates RGB, depth, and edge cues into a joint feature representation. By defining a contrastive loss that pulls rendered features toward similar style features while pushing them away from dissimilar ones, the model ensures that both the surface appearance and the structural form of the 3D scene are aligned with the target style.

Experimental results across various datasets, including LLFF and MipNeRF-360, demonstrate that the proposed method produces more faithful and structurally consistent stylizations compared to existing 3DGS-based approaches. By actively sculpting the geometry rather than just applying a texture, the framework achieves a higher degree of artistic expression. This work is significant because it moves 3D style transfer beyond simple image-to-image mapping, enabling the creation of stylized 3D environments that maintain geometric coherence across novel viewpoints.