3D Multi-scene Stylization Based on Conditional Neural Radiance Fields

Neural Radiation Field (NeRF) is a scene model capable of achieving high-quality view synthesis, optimized for each specific scene. In this paper, we propose a conditional neural radiation field based on multi-resolution hash coding, enabling high-quality synthesis of new views across multiple scenes. By employing multi-resolution hashing to encode 3D positional information, the multilayer perceptron is lightened, thereby reducing memory consumption. The model introduces two latent encodings: shape encoding and appearance encoding, which enhance our model's performance in new view synthesis and scene interpolation. Furthermore, after achieving a robust geometric reconstruction of multiple scenes, we fix the information affecting the scene geometry and utilize a hypernet to predict the parameters of the multilayer perceptron responsible for scene appearance information. This approach facilitates a generalized style transfer across multiple scenes while maintaining the three-dimensional consistency of the scenes.