Directly from Alpha to Omega: Controllable End-to-End Vector Floor Plan Generation

TL;DR

CE2EPlan is introduced, a controllable end-to-end topology- and geometryenhanced diffusion model that enables the model to learn how to design floor plans directly from data, capturing a wide range of solution paths from input boundaries to complete layouts.

Abstract

Automated floor plan generation aims to create residential layouts by arranging rooms within a given boundary, balancing topological, geometric, and aesthetic considerations. The existing methods typically use a multi-step pipeline with intermediate representations to decompose the prediction process into several sub-tasks, limiting model flexibility and imposing predefined solution paths. This often results in unreasonable outputs when applied to data unsuitable for these predefined paths, making it challenging for these methods to match human designers, who do not restrict themselves to a specific set of design workflows. To address these limitations, we introduce CE2EPlan, a controllable end-to-end topology- and geometry-enhanced diffusion model that removes restrictions on the generative process of AI design tools. Instead, it enables the model to learn how to design floor plans directly from data, capturing a wide range of solution paths from input boundaries to complete layouts. Extensive experiments demonstrate that our method surpasses all existing approaches using the multi-step pipeline, delivering higher-quality results with enhanced user control and greater diversity in output, bringing AI design tools closer to the versatility of human designers.

Figures (11)

• Figure 1: Given a boundary as input, unlike previous methods (a) that decompose the design process into several sub-tasks with fixed solution paths ($\alpha\rightarrow\epsilon\rightarrow\lambda\rightarrow\omega$), our CE2EPlan (b) learns floor plan creation directly from data without predefined pathways. It captures diverse design paths from the input boundary to the output floor plan ($\alpha\rightarrow\omega$) and delivers higher-quality results with enhanced user control and output diversity.
• Figure 2: Overview of our CE2EPlan, which adds noise $\epsilon$ to the ground truth data $x_0$ and learns to reverse this process. At each time step $t$, a multi-condition masking mechanism is first used to obtain $M(x_t)$ to support multiple user inputs. Then, the input boundary $c$, $M(x_t)$, & $t$ are fed into the GATransformer to capture topological properties within the floor plan and infer the corresponding noise $\tilde{\epsilon}$ and $\tilde{x}_0$. The predicted $\tilde{x}_0$ is further refined with three alignment losses to enhance geometric consistency.
• Figure 3: Qualitative comparison on the quality of the generated results between our CE2EPlan and the state-of-the-art methods ($\text{iPLAN}_{\text{I}}$, $\text{DiffPlanner}_{\text{I}}$, Graph2Plan, WallPlan, ActFloor-GAN, & RPLAN) for floor plan generation only from boundary ($\text{Mode}_{\text{auto}}$). The flawed design is highlighted in the red box.
• Figure 4: The controllability of our CE2EPlan for the same input boundary across four user interaction modes, i.e., without user input ($\text{Mode}_{\text{auto}}$), using room type input ($\text{Mode}_{\text{t}}$), using both room type & location input ($\text{Mode}_{\text{t}\&\text{l}}$), and partial input (highlighted with red boxes, $\text{Mode}_{\text{part}}$).
• Figure 5: The output diversity of our CE2EPlan for a single input boundary across various user interaction modes, i.e., without user input ($\text{Mode}_{\text{auto}}$), using room type input ($\text{Mode}_{\text{t}}$), using both room type & location input ($\text{Mode}_{\text{t}\&\text{l}}$), and partial input (highlighted with red boxes) with varying proportions of target information ($25\%~\text{Mode}_{\text{part}}$, $50\%~\text{Mode}_{\text{part}}$, & $75\%~\text{Mode}_{\text{part}}$).