HouseDiffusion: Vector Floorplan Generation via a Diffusion Model with Discrete and Continuous Denoising

TL;DR

HouseDiffusion introduces a diffusion-based Transformer model for vector-floorplan generation that accurately denoises 2D coordinates, capturing complex geometric relationships and outperforming existing methods on the RPLAN dataset.

Contribution

The paper proposes a novel diffusion model with discrete and continuous denoising for vector-floorplan generation, enabling precise geometric control and non-Manhattan structure modeling.

Findings

Significant improvements over state-of-the-art metrics

Capable of generating non-Manhattan structures

Controls the exact number of corners per room

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.