Mitigating Domain Shift in Conditioned Floor Plan Generation: Synthetic Pre-training for Data-Efficient Adaptation

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Architecture & Urban Planning · Depth: Expert, extended

Summary

A study by Adrien Bernhardt from Homiwoo and École Polytechnique addresses the significant domain shift problem in conditioned floor plan generation. It evaluates two prominent generative paradigms, arrangement-based Flow Matching (DPFM [6]) and constraint-based Diffusion (Gueze et al. [5]), across RPLAN, MagicPlan, and Swiss Dwellings datasets. Findings reveal up to an order of magnitude performance degradation when models are transferred across domains. To counter this, the research introduces a procedural method to create a large-scale synthetic training dataset of 135k scenes. This synthetic data enforces physical constraints like non-overlapping rooms and valid door placement, but intentionally uses highly irregular spatial arrangements and aggressive geometric perturbations, resulting in 8.83 corners per room on average. Pre-training on this synthetic data significantly improves zero-shot cross-domain performance, even surpassing in-domain training on MagicPlan for DPFM. It also provides a highly effective initialization for fine-tuning, accelerating adaptation and outperforming real-world baselines by up to 40% in low-data regimes (e.g., 1k samples).

Key takeaway

For Machine Learning Engineers deploying floor plan generative models across diverse architectural domains, you should adopt synthetic pre-training. This approach significantly reduces performance degradation from domain shift and accelerates adaptation, requiring fewer real-world samples for fine-tuning. By prioritizing geometrically diverse, rule-compliant synthetic data over architecturally realistic but domain-specific data, you can achieve robust model initialization and competitive performance even in low-data regimes. This strategy minimizes costly large-scale dataset acquisition.

Key insights

Synthetic pre-training with geometrically implausible but rule-compliant data significantly mitigates domain shift in floor plan generation.

Principles

• Domain shift severely degrades floor plan generative models.
• Decouple spatial assembly rules from architectural style.
• Geometric diversity is critical for generalization.

Method

A procedural pipeline generates 135k synthetic floor plans by sampling room shapes, applying aggressive geometric perturbations (aspect ratio distortion, bumps, scale variance), and densely packing them with enforced physical constraints (non-overlap, valid doors, graph consistency).

In practice

• Pre-train generative models on diverse synthetic data.
• Fine-tune from synthetic prior for low-data adaptation.
• Maximize geometric diversity in synthetic data generation.

Topics

• Floor Plan Generation
• Domain Shift Mitigation
• Synthetic Data Pre-training
• Generative Models
• Data-Efficient Adaptation
• Geometric Augmentation

Code references

• facebookresearch/aria_data_tools

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• The Sequence Knowledge # 780: Synthetic Data for Image Models — Synthetic data generation is crucial for robust vision systems, overcoming real-world data limitations through controlled pixel creation.
• Generative Floor Plan Design with LLMs via Reinforcement Learning with Verifiable Rewards — LLMs can generate highly constrained, structured floor plans by combining supervised fine-tuning with reinforcement learning and verifiable rewards.
• Recursive Learning Without Collapse: A Weighting-Based Stabilization Framework — Optimal weighting in recursive generative model training prevents model collapse by balancing real and synthetic data.
• SynSur: An end-to-end generative pipeline for synthetic industrial surface defect generation and detection — Synthetic data generation can augment scarce real datasets for industrial defect detection, improving model performance.
• How Can We Synthesize High-Quality Pretraining Data? A Systematic Study of Prompt Design, Generator Model, and Source Data — Structured output formats and careful source data selection are key to high-quality synthetic pretraining data.
• Beyond Real Data: Synthetic Data through the Lens of Regularization — An optimal synthetic-to-real data ratio exists, minimizing generalization error based on distribution similarity.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.