TL;DR
This paper introduces GSDiff, a novel framework for automatically generating detailed architectural floorplans as vector graphics. By representing floorplans as structural graphs and employing a diffusion model for node generation and a Transformer for edge prediction, GSDiff overcomes limitations of previous methods, producing more geometrically consistent and semantically rich designs.
What Problem Does This Paper Address?
Designing architectural floorplans is a crucial but time-consuming task. While automated methods exist, they often struggle with generating geometrically accurate and semantically meaningful designs. Existing approaches, whether rule-based or learning-based, tend to produce floorplans with misalignments, overlaps, or gaps, requiring extensive manual post-processing to become usable vector graphics. Furthermore, many learning-based methods generate raster images rather than vector formats, limiting their direct applicability. The challenge is to generate high-quality, semantically rich, and geometrically sound vector floorplans directly, with the ability to handle both unconstrained and constrained generation.
Methods
GSDiff tackles floorplan generation by framing it as a structural graph generation problem. A floorplan is represented as a graph where nodes are wall junctions and edges are wall segments. The process is decoupled into two main stages:
- Node Generation: A diffusion model is used to generate the wall junction nodes. To improve geometric consistency, GSDiff introduces a novel alignment loss that optimizes node alignment in a mixed-base representation, balancing the precision of binary representations with the leniency of real-valued ones.
- Edge Prediction: A Transformer-based model predicts the wall segments (edges) between the generated nodes. To enhance the model's understanding of geometric relationships, an edge perception enhancement strategy is employed. This involves adding random interpolation points on edges and training the model to predict these points, thereby improving its ability to infer intermediate edge structures.
The framework also incorporates geometry enhancement strategies for both node and edge generation. For constrained generation, GSDiff uses specialized encoders for boundary and topology constraints, which guide the diffusion and Transformer models via cross-attention.
Datasets
The primary dataset used for training and evaluation is the RPLAN dataset, which contains over 80,000 residential floorplans with dense annotations. The paper also evaluates GSDiff on the LIFULL dataset, which provides vectorized floorplans with 12 room categories, though it notes the annotation quality of LIFULL is poorer.
Key Insights and Ideas
- Structural Graph Representation is Key: Decomposing floorplan generation into node (junction) and edge (segment) prediction within a structural graph framework simplifies the complex task and allows for better control over geometric and semantic aspects.
- Geometry-Enhanced Diffusion: Simply using diffusion models for node generation isn't enough. The proposed alignment loss, which mixes different radix representations, is crucial for achieving precise geometric consistency in wall junctions.
- Edge Perception Matters: Predicting edges is not just about connecting nodes; understanding the "shape" of the wall segment is important. The edge perception enhancement strategy, by predicting interpolation points, significantly improves the accuracy and plausibility of wall segments.
- Bridging Raster and Vector: While many recent methods focus on raster image generation, GSDiff directly generates vector floorplans, offering a more practical output for architectural design.
- Flexibility in Constraints: The framework's ability to incorporate various constraints (boundary, topology) through dedicated encoders demonstrates its versatility for different design needs.
Key Results
- GSDiff outperforms state-of-the-art methods like Graph2Plan and WallPlan in both unconstrained and boundary-constrained generation, achieving better visual appeal, geometric consistency, and adherence to constraints.
- For topology-constrained generation, GSDiff shows superior performance compared to House-GAN++ and HouseDiffusion, exhibiting greater diversity, better constraint satisfaction, and higher visual quality.
- Quantitative evaluations show GSDiff achieving the lowest FID (4.83) and KID (2.84) in ablation studies, indicating high visual fidelity and similarity to real data distributions.
- The method generates floorplans significantly faster than previous methods, with unconstrained generation taking approximately 0.17 seconds per floorplan.
Limitations and Future Work
- Data Dependency: GSDiff relies on high-quality training data for learning complex node and edge dependencies.
- Scalability: The framework faces scalability challenges for larger and more complex projects.
- Future Directions: The authors suggest exploring semi-supervised learning to reduce data dependency, enhancing constraint handling for more complex scenarios, and improving scalability to broaden the framework's applicability. They also aim to extend the method to datasets with multi-apartment layouts.
Summary updated