Techart Conference

Sukhun Ko 

■ Abstract

3D Gaussian Splatting has emerged as an efficient representation for real-time radiance field rendering, enabling high-quality 3D scene reconstruction and rendering. Recent advances extend this paradigm to 3D scene stylization, aiming to transfer artistic styles into spatially consistent representations. However, existing approaches remain largely limited to color-dominant transformations, lacking the ability to preserve fine-grained texture details or to explicitly model spatial structure such as foreground and background separation. As a result, they offer limited controllability and often fail to produce perceptually coherent stylization.

To address these challenges, we propose a controllable 3D scene stylization framework that incorporates depth-aware texture modulation and region-disentangled multi-reference transfer within the Gaussian Splatting pipeline. Specifically, we leverage depth estimation to assign scale-adaptive textures, allowing foreground regions to exhibit larger and more perceptually salient patterns while maintaining global consistency. In addition, we introduce segmentation-based decomposition to separate foreground and background, enabling region-wise multi-reference stylization with explicit control.

Our framework enhances both structural coherence and stylistic expressiveness in 3D stylization. Experimental results on the LLFF dataset demonstrate improved foreground detail preservation and more realistic texture projection compared to prior methods. Notably, our approach effectively captures texture-centric styles such as ink-wash painting, producing more natural brush stroke patterns and tonal variations.

These results highlight the importance of spatial awareness and region-level control in 3D stylization, and suggest a practical direction toward controllable and perceptually grounded 3D style transfer for applications in media art, XR, and immersive environments.

■ Bio