The paper introduces ‘FacadeNet’, a deep learning approach for synthesizing images of building facades from various viewpoints. The method uses a conditional Generative Adversarial Network (cGAN) and a novel selective editing module. This module uses image embeddings from a pretrained vision transformer to precisely modify view-dependent elements like windows and doors while maintaining the structure of view-independent components like walls. Compared to traditional methods, FacadeNet provides better control over the generation process and has demonstrated superior performance on building facade generation tasks.
Publication date: 2 Nov 2023
Project Page: https://arxiv.org/abs/2311.01240v1
Paper: https://arxiv.org/pdf/2311.01240
