StyleGaussian: Instant 3D Style Transfer with Gaussian Splatting

Kunhao Liu et al. introduce StyleGaussian, an innovative 3D style transfer method that uses 3D Gaussian Splatting to achieve real-time performance and consistency across different views. The approach consists of embedding 2D VGG scene features into reconstructed 3D Gaussians, transforming and decoding features, and then applying a K-nearest-neighbor-based 3D CNN. The result is a system that offers instant 3D stylization while preserving real-time rendering and multi-view consistency.

• Explore the project page
• View the paper
• Download the PDF

Highlights:

• Rapid 3D Stylization: Allowing styles from 2D images to be transferred to 3D scenes immediately.
• Real-Time Rendering: Achieving consistent and real-time rendering without sacrificing multi-view consistency.
• Innovative Techniques: The use of embedding and 3D CNN decoding distinguishes this work from conventional methods.

Reflections: The ability to perform style transfers in seconds without compromising on rendering speed or visual consistency is revolutionary. This kind of technology has significant potential not just for graphics designers but also for virtual and augmented reality applications, where rendering speed is crucial.