A Unified Cartoon Rendering Framework for 3-D Gaussian Splatting With Proxy Depth and Stylized Camera Modeling
This paper presents an integrated rendering framework that enables stable cartoon-style rendering in 3D Gaussian Splatting (3DGS) scenes without explicit mesh reconstruction, while reinterpreting camera parameters as expressive controls for cartoon-style cinematography. Because 3DGS generates images through the accumulation of numerous Gaussian primitives, explicit surface depth or a single coherent geometric structure is not directly defined, which often leads to distant noise and visual instability when applying contour enhancement or camera effects. To address this limitation, we utilize the sparse point cloud obtained during the Structure-from-Motion (SfM) stage as proxy geometry to construct an auxiliary depth signal, referred to as proxy depth, which serves as a common cue for both distance-aware cartoon stylization and stylized camera control. The proxy depth is linearized and conservatively refined to reliably separate foreground and background regions, and distance-aware contour attenuation effectively suppresses undesired high-frequency edges in distant regions. Furthermore, this work redefines camera parameters such as focal length, depth of field, and projection not as constraints for photorealistic reproduction, but as expressive control variables for cartoon-style rendering, while maintaining compatibility with the cumulative rendering characteristics of 3DGS. This allows the framework to maintain consistent spatial emphasis and visual staging across viewpoint changes and varying scene compositions. In addition, a stabilization strategy is introduced during training by removing Gaussians with negligible screen-space contributions and constraining the total number of primitives, thereby improving the visual stability of both contour rendering and camera exaggeration. Experimental results demonstrate that the proposed framework achieves stable cartoon contour rendering and expressive camera motion simultaneously in the tested indoor and outdoor 3DGS scenes, while maintaining real-time performance under our experimental setup.