Enhanced 3D Style Transfer of Neural Radiance Fields via Generalized Gaussian Distribution Transformation
Jiayin Yang, Jialun Zhou, Zeyu Wang, Ning Wang · 2024
The endeavor of 3D style transfer is to create styled novel perspectives of a 3D environment, ensuring consistency among multiple views. However, the majority of current works encounter problems such as color distortion or the loss of details in the styled outcomes, which are caused by the disregard for the holistic statistics of the reference image color style. To address these issues, we propose a method called ggdT-StyleRF based on the optimal transport theory of Generalized Gaussian Distribution (GGD). By integrating a GGD transformation step into the StyleRF workflow, we effectively combine color processing with 3D scene style transfer. The GGD transformation method was selected for its demonstrated flexibility and adaptability in statistical modeling, as well as its successful application in 2D image style transfer. Experimental results demonstrate that ggdT-StyleRF significantly enhances the transmission of the reference image color style while preserving the source image's texture structure. Notably, ggdT-StyleRF achieves substantial improvements in Frechet inception distance and Bhattacharyya distance, indicating a significant improvement in color style transfer without compromising geometric fidelity.