Skip to main content
QUICK REVIEW

[Paper Review] FloorplanVLM: A Vision-Language Model for Floorplan Vectorization

Yuanqing Liu, Ziming Yang|arXiv (Cornell University)|Feb 6, 2026
3D Shape Modeling and AnalysisEngineering0 citations
TL;DR

FloorplanVLM reframes floorplan vectorization as image-conditioned sequence generation that outputs structured JSON, achieving engineering-grade, topologically valid floorplans, including non-Manhattan geometries, via a three-stage training and scalable data engine.

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 $ extbf{92.52%}$ external-wall IoU and robust generalization across non-Manhattan architectures.

Motivation & Objective

  • Motivate engineering-grade floorplan vectorization by enforcing strict topological consistency beyond pixel-based methods.
  • Propose an end-to-end sequence modeling framework that outputs JSON representations of walls, openings, and rooms.
  • Create a scalable data engine to balance geometric diversity and pixel-aligned precision for training.
  • Develop a progressive training pipeline combining supervised fine-tuning and reinforcement learning for geometric alignment.
  • Establish FPBench-2K as an open benchmark to standardize evaluation on complex floorplans.

Proposed method

  • Formulate floorplan vectorization as image-conditioned token sequence generation into a hierarchical JSON representation.
  • Introduce a dependency-ordered serialization that first defines a wall skeleton and then rooms referencing those walls.
  • Build Floorplan-2M via structure-aware clustering and Floorplan-HQ-300K for pixel-aligned precision to train with SFT and GRPO.
  • Apply three-stage progressive training: Stage 1-2 supervised fine-tuning for syntactic grounding and quality annealing, Stage 3 Group Relative Policy Optimization to optimize geometric alignment.
  • Use a hierarchical reward function in GRPO combining validity, external geometry IoU, and internal structure to guide learning.
  • Utilize a token-efficient JSON schema with coordinate normalization and 1,391 semantic tokens to compress sequence length.

Experimental results

Research questions

  • RQ1Can FloorplanVLM reliably produce watertight, topologically valid floorplans from raster inputs?
  • RQ2How does the proposed data engine affect diversity and precision of learned geometry across Manhattan and non-Manhattan layouts?
  • RQ3Does progressive training with GRPO improve geometric fidelity beyond standard supervised fine-tuning?
  • RQ4Is the JSON-based representation effective for encoding complex architectural topology compared to alternative formats?

Key findings

  • FloorplanVLM achieves 92.52% external-wall IoU on FPBench-2K, demonstrating strong topological fidelity.
  • Overall, FloorplanVLM attains 96.10% validity, 0.9252 IoU_ext, 0.8920 IoU_room, 0.8249 F1_room, and 0.7333 F1_op on FPBench-2K.
  • The Manhattan subset attains 97.02% validity, 0.9459 IoU_ext, 0.9089 IoU_room, 0.8385 F1_room, and 0.7739 F1_op; the Non-Manhattan subset attains 95.10% validity, 0.9027 IoU_ext, 0.8738 IoU_room, 0.8101 F1_room, and 0.6894 F1_op.
  • A three-stage training regime (SFT plus GRPO) substantially improves geometric alignment and reduces topological errors compared to SFT alone.
  • GRPO provides a significant boost in validity and Ext-IoU, validating the importance of non-differentiable geometry-alignment during training.
  • The JSON-based serialization with structure-first ordering outperforms a Python DSL, driven by pre-training alignment to JSON syntax.

Read this paper

Open the full text in the Nubint reader and ask about any passage as you read.

Read this paper on Nubint

This review was generated by AI from the original paper. Please check the paper before citing it.