HouseDiffusion: Vector Floorplan Generation via a Diffusion Model with Discrete and Continuous Denoising (2211.13287v1)

Abstract: 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.

• The paper introduces HouseDiffusion, a novel diffusion model that integrates discrete and continuous denoising to generate high-quality vector floorplans.
• It leverages a 1D polygonal loop representation and tailored Transformer attention mechanisms to ensure realistic, constraint-compliant architectural geometries.
• Evaluation on the RPLAN dataset reveals a 67% improvement in diversity and a 32% enhancement in compatibility compared to prior approaches.

An Overview of HouseDiffusion: Vector Floorplan Generation Using a Diffusion Model

The paper "HouseDiffusion: Vector Floorplan Generation via a Diffusion Model with Discrete and Continuous Denoising" presents a novel approach to generating vector floorplans by leveraging diffusion models in a manner specifically designed for architectural geometries. Despite the existing advancements in floorplan generation, the authors identify significant shortcomings in the current state-of-the-art approaches, specifically in terms of compliance with input constraints and the ability to generate realistic, varied samples. HouseDiffusion addresses these issues by employing a diffusion model where room and door coordinates are refined through iterative denoising, seamlessly integrating discrete and continuous denoising processes.

The primary innovation of HouseDiffusion lies in its representation of floorplans as 1D polygonal loops, a structure that aligns well with architectural entities such as rooms and doors, allowing it to circumvent the complications associated with rasterized approaches. The authors introduce a system where each floorplan is generated from a bubble diagram representing graph constraints, with nodes corresponding to rooms and edges corresponding to door connections.

The architecture of HouseDiffusion is based on the Transformer, incorporating component-wise self-attention, global self-attention, and relational cross-attention. These tailored attention mechanisms are designed to explicitly handle different relational aspects of the vector geometry, ensuring the generation of structurally sound and realistic floorplans. This system efficiently supports non-Manhattan geometries and allows precise control over room shapes by specifying the number of corners.

The performance of HouseDiffusion was rigorously evaluated using the RPLAN dataset, with results showcasing remarkable improvements in diversity, compatibility, and realism metrics over previous approaches like House-GAN++. These advancements are significant, with HouseDiffusion reportedly achieving a 67% improvement in diversity and a 32% improvement in compatibility metrics over current benchmarks. The system consistently demonstrated its capability to generate high-quality, diverse floorplans that adhere closely to input constraints, even in cases requiring non-standard geometries.

From a theoretical standpoint, this paper extends the application domain of diffusion models into structured geometry generation, contributing a new method that balances continuous coordinate adjustment with discrete state precision. Practically, it holds promise for democratizing architectural design, making sophisticated tools accessible even in environments lacking professional architectural services.

Speculatively, HouseDiffusion invites further exploration into large-scale architectural generation and could eventually extend to incorporate broader urban planning tasks. Future research might explore integration with more complex building regulations or multi-floor architectural layouts, thereby amplifying its utility across architecture and design disciplines. As machine learning continues to push boundaries in graphical and structural tasks, HouseDiffusion sets a precedent for approaches that fuse geometric fidelity with generative creativity.