Eliminating Rasterization: Direct Vector Floor Plan Generation with DiffPlanner

2025-08-19

AI Summary

Generating planar floor layouts under boundary constraints requires joint modeling of room topology and geometric arrangement; however, existing learning-based approaches rely on a “vector→raster→vector” pipeline, causing quantization errors and scale-dependent distortions. This paper proposes DiffPlanner—the first end-to-end vector-space generative framework for high-fidelity, editable floor plan synthesis. Its core contributions are: (1) a conditional diffusion model based on Transformer architecture, enabling direct sequential modeling of vectorized layout primitives (e.g., walls, doors); and (2) a learnable alignment mechanism that explicitly emulates iterative design refinement, enhancing controllability and perceptual realism. Extensive experiments demonstrate that DiffPlanner significantly outperforms state-of-the-art methods in both quantitative metrics (e.g., FID, topology accuracy) and human evaluation, particularly excelling in early-stage creative design where high fidelity and post-generation editability are critical.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Computer Vision: Diffusion Models for Vision
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

The boundary-constrained floor plan generation problem aims to generate the topological and geometric properties of a set of rooms within a given boundary. Recently, learning-based methods have made significant progress in generating realistic floor plans. However, these methods involve a workflow of converting vector data into raster images, using image-based generative models, and then converting the results back into vector data. This process is complex and redundant, often resulting in information loss. Raster images, unlike vector data, cannot scale without losing detail and precision. To address these issues, we propose a novel deep learning framework called DiffPlanner for boundary-constrained floor plan generation, which operates entirely in vector space. Our framework is a Transformer-based conditional diffusion model that integrates an alignment mechanism in training, aligning the optimization trajectory of the model with the iterative design processes of designers. This enables our model to handle complex vector data, better fit the distribution of the predicted targets, accomplish the challenging task of floor plan layout design, and achieve user-controllable generation. We conduct quantitative comparisons, qualitative evaluations, ablation experiments, and perceptual studies to evaluate our method. Extensive experiments demonstrate that DiffPlanner surpasses existing state-of-the-art methods in generating floor plans and bubble diagrams in the creative stages, offering more controllability to users and producing higher-quality results that closely match the ground truths.

Problem

• Generates vector floor plans without rasterization
• Handles complex vector data in boundary constraints
• Enables user-controllable layout design generation

Innovation

• Transformer-based conditional diffusion model
• Operates entirely in vector space
• Alignment mechanism matching design processes