BIMFusion: Texturing Building Information Models Using Diffusion Generative Methods

Chao Deng, Han Huang, Ziyi Ke, Yulun Wu, Ge Gao, Ming Gu

Abstract: Models created by building information modeling (BIM) represent scenes with rich semantics but are usually untextured. The absence of authentic textures impedes BIM-related tasks and BIM’s downstream applications such as interior design. This paper aims to style a BIM indoor scene with the help of diffusion models, a family of generative algorithms capable of synthesizing high-quality images. The proposed pipeline BIMFusion classifies elements in an indoor scene into structural and furniture elements. Texturing structural elements is achieved by first synthesizing tileable texture candidates and then selecting the optimal candidates with an automatic rating method. For furniture elements, this paper develops a geometry-aware method to texture furniture within a scene, based on a multiview 3D texture algorithm but adapted to better fit the characteristics of BIM models. Qualitative and quantitative results show that the generated textures are visually satisfactory, strictly aligned with the model’s geometry and semantics, and comply with the user’s intended styles.

External IDs:dblp:journals/jccee/DengHKWGG26