Constrained Layout Generation with Factor Graphs

TL;DR

This work tackles constrained floorplan layout generation by encoding higher-order spatial constraints as a factor graph, where each room is represented by four bounding-box variables $(x_{min}^i, x_{max}^i, y_{min}^i, y_{max}^i)$. A Floor-Plan Factor Graph Neural Network (FP-FGNN) performs neural message passing over variable and factor nodes—including box, relation, boundary, and complete factors—to infer bounding boxes that satisfy user-defined and structural constraints, then converts these boxes into a floorplan image via a CRN-based refinement step. FP-FGNN demonstrates substantial gains over Graph2Plan and HPGM in box-level IOU (roughly a 20% improvement) and strong fidelity to constraints, while supporting rapid iterative design (inference ≈0.023s) and integration into generative pipelines for diversity (via boundary-based retrieval and favorable FID scores). The approach highlights the value of higher-order, domain-knowledge-driven constraints in practice, enabling accurate, interactive AI-guided floorplan design. Future work includes extending FP-FGNN to incorporate language-described requirements for even more expressive, natural human-in-the-loop interaction.

Abstract

This paper addresses the challenge of object-centric layout generation under spatial constraints, seen in multiple domains including floorplan design process. The design process typically involves specifying a set of spatial constraints that include object attributes like size and inter-object relations such as relative positioning. Existing works, which typically represent objects as single nodes, lack the granularity to accurately model complex interactions between objects. For instance, often only certain parts of an object, like a room's right wall, interact with adjacent objects. To address this gap, we introduce a factor graph based approach with four latent variable nodes for each room, and a factor node for each constraint. The factor nodes represent dependencies among the variables to which they are connected, effectively capturing constraints that are potentially of a higher order. We then develop message-passing on the bipartite graph, forming a factor graph neural network that is trained to produce a floorplan that aligns with the desired requirements. Our approach is simple and generates layouts faithful to the user requirements, demonstrated by a large improvement in IOU scores over existing methods. Additionally, our approach, being inferential and accurate, is well-suited to the practical human-in-the-loop design process where specifications evolve iteratively, offering a practical and powerful tool for AI-guided design.

Figures (11)

• Figure 1: Given a building boundary and a graph that encodes user requirements and room constraints, we first transform the inputs into a factor graph that can model higher-order constraints, and then produce the floorplan based on the factor graph model.
• Figure 2: Illustration of the proposed factor graph model for floorplan design. Each room is represented with four bounding-box variables. Factors connect the variables based on input constraints and domain knowledge.$f_{box}^i$ connects four variables of room $i$. $f_{rel}^{s,o}$ connects only relevant subset of variables between rooms $s$ and $o$. $f_{boundary}^k$ represents $k^{th}$ corner-point and connects to all the variables. Message passing on this factor graph helps learn better room coordinates which are then used to produce the layout image.
• Figure 3: Visualization of layout predictions of Graph2Plan and FP-FGNN without postprocessing. FP-FGNN is able to produce more accurate predictions of room dimensions and the shape of walls, resulting in layouts that are more visually coherent and well-defined.
• Figure 4: Illustration of iterative design with user iteration. Starting from the same arbitrary partial layout, the interaction process results in three different final layout. The room sizes are automatically adjusted for alignment with boundaries and adjacent rooms.
• Figure 5: Macro IOU on specified rooms under partial information.