StyleNGP: A fast way to stylize 3D scene
Weikang Hu, Xianhua Duan, Shuai Shen · 2024
3D scene stylization aims to stylize the implicit representation with an arbitrary reference image. Combining existing approaches for novel view synthesis and 2D stylization directly often causes jittering artifacts due to the lack of cross-view consistency. SNeRF addressed this problem by alternating the NeRF and stylization optimization steps, which is memory efficient but time-consuming. In this work, an enhanced training approach for stylizing 3D scenes is introduced using NGP and a consist style transfer method. Another problem is that 2D style transfer approaches can not preserve depth information well. To alleviate it, a depth loss function has been integrated into the approach, which utilizes the pseudo depth of NGP as ground truth. Experimental results demonstrate that the enhanced approach can transfer more details to 3D scenes while maintaining cross-view consistency, within a shorter training time.