Constrained Layout Generation with Factor Graphs

This paper addresses the challenge of object-centric layout generation underspatial constraints, seen in multiple domains including floorplan designprocess. The design process typically involves specifying a set of spatialconstraints that include object attributes like size and inter-object relationssuch as relative positioning. Existing works, which typically represent objectsas single nodes, lack the granularity to accurately model complex interactionsbetween objects. For instance, often only certain parts of an object, like aroom's right wall, interact with adjacent objects. To address this gap, weintroduce a factor graph based approach with four latent variable nodes foreach room, and a factor node for each constraint. The factor nodes representdependencies among the variables to which they are connected, effectivelycapturing constraints that are potentially of a higher order. We then developmessage-passing on the bipartite graph, forming a factor graph neural networkthat is trained to produce a floorplan that aligns with the desiredrequirements. Our approach is simple and generates layouts faithful to the userrequirements, demonstrated by a large improvement in IOU scores over existingmethods. Additionally, our approach, being inferential and accurate, iswell-suited to the practical human-in-the-loop design process wherespecifications evolve iteratively, offering a practical and powerful tool forAI-guided design.