PanoGS: Panorama-based 3D Scene Stylization with Style and Geometry Consistency.

Abstract

3D scene stylization offers enhanced immersion and visual coherence for 3D experiences in digital content creation and augmented reality applications. Existing approaches frequently struggle to balance style consistency across different viewpoints, geometry consistency as reference, and zero-shot generalization to unseen styles. To overcome these problems, we present PanoGS, a panorama-based 3D scene stylization framework incorporating two innovative designs: Circular Consistent Stylization and Progressive Propagation Refinement. Circular Consistent Stylization imposes style priors, geometric structure, and circular denoising on a Stable Diffusion-based stylization, ensuring both faithful style transfer from reference images and preservation of original scene geometry with seamless circular continuity. Progressive Propagation Refinement extends single-viewpoint panorama stylization to multi-viewpoint panoramas through color propagation and novel-view refinement, effectively addressing the stylization of occluded regions under sparse viewpoints. Experiments demonstrate that PanoGS outperforms existing methods in quantitative metrics, with an average improvement of 9.23% on CLIP and 29.43% and 28.58% on short-range LPIPS and RMSE. Furthermore, user studies indicate strong preference rates for our method regarding geometric (75.00%), stylistic (66.37%), and semantic consistency (74.11%). Overall, our approach achieves high-quality 3D stylization with superior style consistency and geometric alignment to input in a zero-shot manner.