FloorplanVLM: A Vision-Language Model for Floorplan Vectorization

TL;DR

FloorplanVLM reframes floorplan vectorization as image-conditioned sequence generation that outputs a structured JSON describing walls, openings, and rooms, enabling topology-aware, engineering-grade vector graphics. It couples a scalable data engine (Floorplan-2M and Floorplan-HQ-300K) with a three-stage training pipeline, culminating in Group Relative Policy Optimization (GRPO) to enforce strict geometric alignment and watertightness. On FPBench-2K, FloorplanVLM achieves an external-wall IoU of 0.9252 and strong generalization to non-Manhattan geometries, validating end-to-end sequence modeling for complex architectural layouts. The work also provides FPBench-2K as an open benchmark and demonstrates how data scale, structure-aware sampling, and RL-based geometric rewards together enable robust, scalable floorplan vectorization beyond traditional heuristics.

Abstract

Converting raster floorplans into engineering-grade vector graphics is challenging due to complex topology and strict geometric constraints. To address this, we present FloorplanVLM, a unified framework that reformulates floorplan vectorization as an image-conditioned sequence modeling task. Unlike pixel-based methods that rely on fragile heuristics or query-based transformers that generate fragmented rooms, our model directly outputs structured JSON sequences representing the global topology. This 'pixels-to-sequence' paradigm enables the precise and holistic constraint satisfaction of complex geometries, such as slanted walls and curved arcs. To support this data-hungry approach, we introduce a scalable data engine: we construct a large-scale dataset (Floorplan-2M) and a high-fidelity subset (Floorplan-HQ-300K) to balance geometric diversity and pixel-level precision. We then employ a progressive training strategy, using Supervised Fine-Tuning (SFT) for structural grounding and quality annealing, followed by Group Relative Policy Optimization (GRPO) for strict geometric alignment. To standardize evaluation on complex layouts, we establish and open-source FPBench-2K. Evaluated on this rigorous benchmark, FloorplanVLM demonstrates exceptional structural validity, achieving $\textbf{92.52%}$ external-wall IoU and robust generalization across non-Manhattan architectures.

Figures (8)

• Figure 1: Comparison of Floorplan Vectorization Paradigms. Unlike pixel-based methods that lead to primitive mismatches or query-based methods that cause topological gaps, FloorplanVLM reframes vectorization as sequence modeling. The model directly outputs structured JSON sequences containing geometric attributes like coordinates and curvature, enabling high-quality floorplan reconstruction.
• Figure 2: Overview of FloorplanVLM Framework. We reformulate floorplan vectorization as a multi-modal sequence generation task using Qwen2.5-VL. (Left) The model accepts a raster floorplan and a prompt, transforming visual data into a discrete token sequence. (Center) A progressive training pipeline bridges the gap between visual perception and geometric logic: Stages 1 & 2 (SFT) establish syntactic correctness and generation quality, while Stage 3 employs Group Relative Policy Optimization (GRPO) to enforce strict geometric alignment (e.g., closed loops). (Right) The output is a structured, hierarchical JSON sequence that deterministically parses into a precise vector floorplan.
• Figure 3: Overview of the Scalable Data Engine. The pipeline illustrates the construction of our hierarchical dataset structure. (Top) Starting from a raw pool of 20M samples, we construct Floorplan-2M via structure-aware clustering to ensure geometric diversity. Subsequently, we curate the high-fidelity Floorplan-HQ-300K subset through a hybrid process combining human recaption and synthetic rendering (for pixel-perfect alignment), serving as the foundation for Stage 2 & 3 training. (Bottom) To enable balanced sampling, we extract dual-view features: Geometry Contour utilizes Fourier Descriptors to capture global boundary invariance, while Layout Topology employs graph embeddings to encode the internal spatial logic of room connections.
• Figure 4: Statistical Distribution of Training Datasets. We compare the geometric type (left) and primitive count (right) distributions between the raw Floorplan-2M(top) and the refined Floorplan-HQ-300K(bottom). Our structure-aware sampling effectively rebalances the dataset, significantly increasing the proportion of non-Manhattan geometries (from 36.3% to 42.7%) while preserving the long-tail distribution of scene complexity.
• Figure 5: Statistical Distribution of FPBench-2K.