Source-linked AI summary
HouseDiffusion: Vector Floorplan Generation via a Diffusion Model with Discrete and Continuous Denoising
Mohammad Amin Shabani, Sepidehsadat Hosseini, Yasutaka Furukawa
TL;DR
Graph-constrained floorplan generation requires vector outputs with precise architectural relationships that raster methods and coordinate regression handle poorly. HouseDiffusion uses polygonal-loop representations and Transformer-based diffusion with continuous and discrete denoising, outperforming House-GAN++ while supporting non-Manhattan structures and exact corner counts. Its reported future work is handling large-scale buildings.
Problem
Existing floorplan generators can violate input constraints, lack variation, require vectorization post-processing, or fail to establish precise geometric relationships.
Method
HouseDiffusion represents rooms and doors as polygonal loops and uses a Transformer diffusion model with continuous noise denoising and discrete coordinate denoising under graph constraints.
Results
HouseDiffusion outperforms House-GAN++ with significant margins on all reported metrics while generating non-Manhattan floorplans and controlling exact corners per room.
Takeaways & Limitations
The approach directly generates vector-graphics floorplans end-to-end without post-processing and establishes a compelling method for structured geometric generation.
Takeaways & Limitations
Future work includes handling large-scale buildings.
Abstract
from arXiv · showhide
The paper presents a novel approach for vector-floorplan generation via a diffusion model, which denoises 2D coordinates of room/door corners with two inference objectives: 1) a single-step noise as the continuous quantity to precisely invert the continuous forward process; and 2) the final 2D coordinate as the discrete quantity to establish geometric incident relationships such as parallelism, orthogonality, and corner-sharing. Our task is graph-conditioned floorplan generation, a common workflow in floorplan design. We represent a floorplan as 1D polygonal loops, each of which corresponds to a room or a door. Our diffusion model employs a Transformer architecture at the core, which controls the attention masks based on the input graph-constraint and directly generates vector-graphics floorplans via a discrete and continuous denoising process. We have evaluated our approach on RPLAN dataset. The proposed approach makes significant improvements in all the metrics against the state-of-the-art with significant margins, while being capable of generating non-Manhattan structures and controlling the exact number of corners per room. A project website with supplementary video and document is here https://aminshabani.github.io/housediffusion.
1. Introduction
Existing floorplan generators struggle with constraint compatibility, variation, vector structure, and precise geometric relationships. HouseDiffusion addresses these challenges with graph-conditioned diffusion that directly generates vector floorplans and improves reported metrics.
- 1. Introduction: The motivation includes making floorplan-design technology more affordable for house buyers because most North American buildings lack dedicated architects.The passage attributes this lack of dedicated architects primarily to cost.
- 1. Introduction: Existing models can violate input constraints, lack variation, or produce outputs that do not resemble floorplans.Raster convolutions support local refinement but lack global reasoning and require post-processing for vectorization.
- 1. Introduction: Direct coordinate regression cannot reliably establish exact architectural relationships such as shared corners and axis-aligned walls.Discrete coordinate encodings avoid regression limitations but suffer severe label imbalance during training.
- 1. Introduction: HouseDiffusion represents rooms and doors as 1D polygonal loops and denoises coordinates with continuous noise and discrete final-coordinate objectives.The continuous branch inverts the forward process, while the discrete branch establishes incident relationships.
- 1. Introduction: The method directly generates vector-graphics floorplans without post-processing, supports non-Manhattan architectures, and improves all reported metrics.Its graph constraint is a bubble diagram whose nodes are rooms and whose edges are door connections.
2. Related Work
Prior work includes raster and vector floorplan generators, graph-theoretic optimization, and diffusion models across several domains. HouseDiffusion differs by processing vector geometry throughout while targeting precise architectural relationships.
- 2. Related Work: House-GAN and House-GAN++ generate graph-constrained floorplans through room segmentation masks, with House-GAN++ iteratively refining layouts.These methods represent rooms through raster masks rather than maintaining vector geometry throughout generation.
- 2. Related Work: Other approaches combine vector generation with raster discrimination or use graph-theoretic and linear optimization techniques for floorplan construction.The vector-generator approach is limited to rectangular shapes.
- 2. Related Work: HouseDiffusion processes vector geometry from start to finish under a bubble-diagram constraint, unlike prior raster-based floorplan workflows.The bubble diagram uses rooms as nodes and door connections as edges.
- 2. Related Work: Diffusion models have been applied to image, language, forecasting, speech, text-to-image, molecular conformation, and 3D-shape generation.These applications motivate diffusion as an emerging general-purpose generative framework.
- 2. Related Work: Floorplan generation is uniquely challenging because continuous coordinate regression rarely yields exact parallelism, orthogonality, or corner-sharing relationships.Related diffusion work therefore explores discrete state spaces or learned embeddings for discrete data.
3. Preliminary
Diffusion models transform data into progressively noisier samples and learn the reverse process that denoises Gaussian noise back toward a data sample. A noise schedule controls the forward corruption over time.
- 3. Preliminary: Diffusion-model training consists of forward noising and reverse denoising processes performed over T steps.The reverse process begins from Gaussian noise and progressively approaches the data sample.
- 3. Preliminary: The forward process creates x_t from x_0 by adding sampled Gaussian noise ϵ ∼ N(0, I).The noise amount depends on the time step t.
- 3. Preliminary: The forward equation mixes x_0 with Gaussian noise using γ_t and sqrt(1−γ_t), producing the noisy sample x_t.The supplied equation passage gives the schedule-weighted combination of clean data and noise.
- 3. Preliminary: γ_t is a noise schedule that gradually changes from 1 to 0.The reverse process starts from a pure Gaussian-noise sample x_T and estimates earlier states step by step.
4. HouseDiffusion
HouseDiffusion encodes graph-constrained floorplans as polygonal loops and uses a Transformer with structured attention to denoise their corner coordinates. Continuous and discrete branches jointly produce geometrically consistent vector outputs.
- 4. HouseDiffusion: The bubble-diagram condition is a graph whose nodes are rooms with types and whose edges denote door connections.Room types and door types are explicitly enumerated in the representation metadata.
- 4. HouseDiffusion: A floorplan is represented as one 1D polygonal loop per room or door, with each loop defined by ordered 2D corner coordinates.This representation requires exact sharing of room corners and walls without gaps or overlaps.
- 4.1. Floorplan Representation: The number of corners N_i is sampled from type-specific training histograms or specified directly to control room-shape complexity.Avoiding a maximum-size representation reduces representation size and training difficulty.
- 4.1. Floorplan Representation: Coordinates are integer values in [0,255], mapped to [−1,1] for continuous Gaussian noising and represented as 8-bit binary values for discrete denoising.The discrete representation supports precise coordinate relationships during reverse inference.
- 4.2. HouseDiffusion Architecture: Corner embeddings combine augmented wall samples, room type, room and corner indices, and time before projection to 512 dimensions.Sampling eight points along the wall to the next corner helps reason about incident relationships along walls.
- 4.2. HouseDiffusion Architecture: The Transformer uses Component-wise Self Attention, Global Self Attention, and Relational Cross Attention with structured masks tied to rooms, all corners, and graph-connected rooms and doors.The three attention types respectively model within-component relations, global corner interactions, and room-door connections.
- 4.2. HouseDiffusion Architecture: The discrete branch converts denoised coordinates to rounded 8-dimensional binary vectors and thresholds them at test time to obtain integer coordinates.It is trained only near the end of denoising, when the final-coordinate inference is accurate; at testing, discrete coordinates are passed forward only for t < 32.
5. Experiments
HouseDiffusion is evaluated against prior graph-constrained floorplan generators on quantitative, qualitative, and ablation studies. It consistently improves benchmark performance while preserving input constraints and supporting controllable geometric generation.
- 5.1. Quantitative Evaluations: 67% higher diversity and 32% higher compatibility than House-GAN++ establish HouseDiffusion’s main quantitative advantage, including significant gains on non-Manhattan RPLAN.House-GAN++ cannot handle non-Manhattan structures, and realism is not evaluated on that benchmark.
- 5.1. Quantitative Evaluations: HouseDiffusion achieves a realism score of −0.19 against ground truth, with participants judging its output equal to ground truth 81% of the time.Its direct network outputs avoid the expensive post-processing and incompatibility filtering used by House-GAN++’s cited best variant.
- 5.2. Qualitative Evaluations: Qualitatively, HouseDiffusion produces higher-quality, more diverse samples with exactly the input rooms and clean wall structures, unlike House-GAN++’s duplicate or missing rooms.The comparison covers both Manhattan and non-Manhattan bubble diagrams.
- 5.3. Ablation Studies: The discrete-and-continuous denoising scheme significantly increases diversity with a small compatibility sacrifice compared with AnalogBits and continuous-only denoising.The authors hypothesize that AnalogBits mainly learns rounding when the diffusion timestep is small.
- 5.3. Ablation Studies: Global self-attention most affects realism and diversity, relational cross-attention most affects compatibility, and component-wise self-attention most affects room-shape quality.Removing CSA causes self-intersections and impossible shapes, while removing RCA corrupts door placements.
- 5.3. Ablation Studies: Corner augmentation improves both diversity and compatibility, while activating the discrete branch during the final 32 iterations provides the best overall performance.Augmented wall points improve analysis of geometric incident relationships, whereas too few or too-early discrete iterations degrade results.
6. Conclusion
HouseDiffusion directly generates vector-graphics floorplans using diffusion and Transformer attention over structured architectural components. It improves all reported metrics over the state of the art while supporting non-Manhattan structures and exact corner counts.
- HouseDiffusion combines discrete and continuous denoising of 2D coordinates with a Transformer whose attentions exploit architectural component relationships.The discrete representation preserves geometric incident relationships among rooms and doors.
- The system significantly improves over the current state of the art across all metrics with large margins.
- HouseDiffusion generates non-Manhattan structures and allows exact specification of the number of corners.The paper identifies large-scale-building handling as future work.