A controllable generative design framework for residential communities with multi-scale architectural representations

Abstract

We present a novel human-in-the-loop framework, CLOD-ReCo, for controllable residential community (ReCo) layout design in the form of multiple levels-of-detail (LODs) for a given construction plot boundary. Unlike other existing end-to-end methods that can only predict a basic 2D raster ReCo plan (LOD0), our approach simulates the design process of architects, which can not only be automated to generate diverse, vector-based, and high-quality 3D ReCo plans (LOD1$\sim$4), but can also interact with the users during the entire generation process, from sketching, including the building numbers and locations to LOD4 including a realistic representation of a group of buildings and their surroundings, making humans and AI co-design the final layout plan. Intensive experiments are conducted to demonstrate the strengths of our approach. The quantitative evaluation, the qualitative comparison, and the subjective evaluation by architects show the ability of our method to generate high-quality and plausible results, which are better than those produced by prior existing methods and comparable to the real-world ReCo plans designed by professional architects. Furthermore, the experiments on the variability of our automated method and user interaction show the ability of our approach to generate diverse results and to interact with users toward co-designing human-centric ReCo plans that meet the requirements of architects.