Skip to main content
Search

# Furniture Layout Generation

1 related articles

Subscribe to updates
  1. arXiv

    Architect-Ant trains on 505 real professional floor plans to produce editable furniture layouts, reaching 94% functional completeness with 0.6% collision and out-of-bounds rates

    The work introduces AntPlan, a curated dataset of 505 real professional residential floor plans with dense annotations across 92 furniture classes and ten room categories, and Architect-Ant, which represents layouts in an editable coordinate DSL, adapts Gemma-4-31B-it via LoRA supervised fine-tuning, then applies GRPO with a room-aware Layout Rule Score (LRS) as final-layout reward; on 109 SceneSmith room shells it reaches 94% mean functional completeness, LRS 8.80, and 0.6% collision and out-of-bounds rates, outperforming the compared baselines while running at about 95.4 s per room, faster than iterative agentic approaches.

You’re all caught up