Source-linked AI summary
Graph2Plan: Learning Floorplan Generation from Layout Graphs
Ruizhen Hu, Zeyu Huang, Yuhan Tang, Oliver van Kaick, Hao Zhang, Hui Huang
TL;DR
Floorplan generation needs to accommodate sparse user preferences beyond a building boundary, including room counts and relationships. Graph2Plan retrieves and edits layout graphs, then generates boundary-conforming floorplans, producing varied and high-quality vectorized results while leaving some constraints unmodeled.
Problem
Existing floorplan inputs may not capture user preferences such as room counts, locations, and connectivity, motivating sparse layout-graph constraints.
Method
Graph2Plan retrieves and adjusts layout graphs from RPLAN, then combines graph and boundary processing to generate raster floorplans and room boxes.
Results
The framework generates varied floorplans from one boundary, adapts to edited layout constraints, and produces high-quality vectorized floorplans.
Takeaways & Limitations
Users can guide floorplan generation at a high level while retaining flexibility to explore multiple designs from the same boundary.
Takeaways & Limitations
The layout graphs omit accessibility, functionality, negative adjacency, boundary-feature constraints, and core integration of alignment and vectorization.
Abstract
from arXiv · showhide
We introduce a learning framework for automated floorplan generation which combines generative modeling using deep neural networks and user-in-the-loop designs to enable human users to provide sparse design constraints. Such constraints are represented by a layout graph. The core component of our learning framework is a deep neural network, Graph2Plan, which converts a layout graph, along with a building boundary, into a floorplan that fulfills both the layout and boundary constraints. Given an input building boundary, we allow a user to specify room counts and other layout constraints, which are used to retrieve a set of floorplans, with their associated layout graphs, from a database. For each retrieved layout graph, along with the input boundary, Graph2Plan first generates a corresponding raster floorplan image, and then a refined set of boxes representing the rooms. Graph2Plan is trained on RPLAN, a large-scale dataset consisting of 80K annotated floorplans. The network is mainly based on convolutional processing over both the layout graph, via a graph neural network (GNN), and the input building boundary, as well as the raster floorplan images, via conventional image convolution.
1 INTRODUCTION
Graph2Plan combines deep generative modeling with user-in-the-loop design so sparse room and adjacency constraints can guide floorplan generation. It retrieves and adjusts layout graphs, then generates raster and vectorized floorplans within a building boundary.
- Graph2Plan converts a layout graph and building boundary into a floorplan satisfying both layout and boundary constraints.
- Users can specify room counts, locations, and adjacencies, which are used to retrieve and refine candidate layout graphs from RPLAN.The retrieved graphs can support multiple suitable floorplans because sparse constraints do not fully specify the final design.
- The framework generates a raster floorplan image and room bounding boxes for each retrieved layout graph, then aligns the boxes into a vectorized floorplan.The network uses graph neural processing for layout graphs and convolutional processing for boundaries and raster images.
- The framework is intended to provide flexible control, ranging from automatic variety generation to constraint-guided floorplan mock-ups.Users can explore multiple floorplans from one boundary and edit layout graphs to guide the results.
2 RELATED WORK
Prior floorplan methods use procedural, optimization, or deep-learning approaches, while Graph2Plan adds layout-graph control to deep floorplan generation. Its retrieval-based graphs incorporate user goals and design principles from training examples.
- Floorplan generation: Floorplan generation has traditionally used building outlines, room constraints, and procedural or optimization methods to propose room layouts.
- Deep learning for layout generation: Deep-learning approaches generate floorplans from building outlines, but some provide limited high-level control over room dimensions, specifications, or adjacencies.Wu et al. predict room locations and walls, while ArchiGAN lets users edit images without high-level room specifications.
- Deep learning for layout generation: Graph2Plan lets users specify room adjacencies and high-level room properties through a layout graph, providing finer control than image-editing interfaces.
- Related graph-based methods: The method retrieves layout graphs from floorplans and trains Graph2Plan end-to-end to produce a floorplan image with associated room boxes.
- Related graph-based methods: Unlike methods assuming rectangular room boundaries, Graph2Plan works with arbitrary rectilinear building boundaries.
3 OVERVIEW
The framework retrieves and interactively adjusts layout graphs before using Graph2Plan to retarget them to a new building boundary. The network predicts room boxes and a raster image to support a valid composed floorplan.
- Initial user input: Users enter a building boundary and optional constraints on room counts, locations, and adjacencies to guide layout retrieval and generation.
- Layout graphs: The system filters and ranks RPLAN layout graphs by user constraints and source-boundary similarity, then lets users edit the retrieved graphs.The resulting candidate graphs guide floorplan generation.
- Floorplan generation: Graph2Plan learns to retarget a retrieved layout graph from its source boundary to the input boundary using design principles in the training data.The retargeting problem is difficult because differing boundaries can create conflicts in room locations, sizes, and shapes.
- Floorplan generation: The network takes a layout graph and building boundary, predicts one bounding box per room, and generates a raster image assigning one room label to each pixel.The raster prediction helps compose boxes that may overlap or be imperfectly aligned.
4 LAYOUT GRAPH RECOMMENDATION
Layout graphs encode room attributes and spatial relations, while boundary retrieval and adjustment align candidate graphs with the user’s building. Users can then edit nodes and edges before generation.
- Layout graph extraction: Each layout-graph node represents a room, and edges connect rooms that are adjacent in the floorplan.
- Layout graph extraction: Room nodes encode type, location on a 5 × 5 grid, and size relative to the building.
- Layout graph extraction: Edges encode spatial relations such as left, right, above, below, diagonal, inside, and outside.
- Layout graph retrieval: More than 80K layout graphs are extracted from 120K RPLAN floorplans and used as varied templates for floorplan design.
- Layout graph retrieval: Graphs are filtered by user constraints and ranked by source-boundary similarity, with the turning function incorporating boundary shape and front-door position.
- Layout graph adjustment: Boundary alignment uses front doors as references and restricts transformations to rotations by k × 90 degrees.The chosen rotation makes the front-door directions differ by less than 45 degrees.
- Layout graph adjustment: Transferred nodes outside the input boundary are moved to the closest available grid cells, while users can further add, delete, or move nodes and adjacency edges.
5 GRAPH-BASED FLOORPLAN GENERATION
Graph2Plan maps a user-constrained layout graph and building boundary to raster and room-box floorplans, then refines and vectorizes the result. Its architecture combines graph and image processing with geometric losses and post-processing for alignment and room ordering.
- Network inputs and outputs: The network encodes a building boundary and layout graph as inputs, producing initial and refined room boxes plus a 128 × 128 floorplan image.Each predicted box is represented by position and size parameters [xi, yi, wi, hi].
- Network architecture: A GNN embeds rooms and their relations, while boundary features are concatenated with room features to predict boxes and guide raster-image generation.Overlapping room features are summed when composing the floorplan image.
- Network architecture: BoxRefineNet uses a CNN feature map, RoI pooling, and room features to predict each room’s refined box position and size.The refinement network processes each predicted box as a region of interest.
- Loss functions: The training objective combines image cross-entropy, box regression, and geometric consistency losses for initial boxes.The geometric loss constrains consistency among boxes and between boxes and the input boundary.
- Loss functions: Geometric terms enforce building coverage, boundary interiority, low box overlap, and agreement with ground-truth boxes.Coverage, interior, mutex, and match terms respectively address these four constraints.
- Room alignment and floorplan vectorization: A final vectorization step uses the raster image to resolve overlapping room labels, align room boundaries, and determine a valid drawing order.The ordering procedure constructs directed constraints between overlapping rooms and removes nodes with outdegree zero; loops are broken by deleting a minimum-outdegree node.
6 RESULTS AND EVALUATION
Graph2Plan generates diverse, constraint-satisfying floorplans across complex boundaries and user edits, with quantitative evaluations showing plausible room boxes and benefits from image guidance, geometric losses, and refinement.
- Qualitative results: Different boundaries, constraints, and front-door locations produce varied room arrangements while preserving specified room counts and relationships.
- Qualitative results: The framework generates reasonable floorplans for complex building boundaries and layout constraints, including graphs with up to eight nodes.
- Graph adjustment: User graph edits change targeted parts of the layout while preserving other regions, and adding rooms updates the generated floorplan.
- Qualitative evaluation: 49.3% of user-study responses judged the generated floorplans at least as plausible as ground truth, compared with 50.7% favoring ground truth.The study included 30 participants and 600 answers.
- Quantitative evaluation: 0.65 average IoU indicates good room-box prediction, while predicted boxes generally match ground-truth location and size.Small differences can occur near room boundaries.
- Ablation study: Image guidance prevents some rooms from disappearing, geometric loss slightly improves box IoU, and refinement significantly improves performance and tightens room boxes.Across post-processing comparisons, IoU rises from 0.54 to 0.66 before post-processing and from 0.75 to 0.80 after it.
7 CONCLUSION AND FUTURE WORK
Graph2Plan enables user-in-the-loop floorplan generation by combining layout-graph constraints with building boundaries, but its constraint types and robustness remain limited. The authors identify extensions toward richer graph manipulations and related design synthesis.
- Graph2Plan lets users specify room types, counts, and adjacencies, then refine retrieved layout graphs to generate varied floorplans from the same boundary.The framework supports fine-tuning results through layout-graph edits.
- Quantitative evaluation shows generated floorplans resemble training examples and tend to follow design principles learned from the dataset.
- When a retrieved floorplan boundary differs substantially from the input, graph adjustment can produce overlapping rooms and unsatisfied relationship constraints.The failure occurs when room nodes become distributed too closely together after adjustment.
- The layout graphs omit accessibility, functionality, interior doors, windows, and some negative-adjacency constraints.These omitted considerations are therefore not captured by the learned network.
- Future work includes guiding furniture synthesis with part layout graphs and supporting more complex layout-graph manipulations.