What they did

The authors designed a framework with three specialized agents: a Generator proposes furniture placements, an Evaluator checks geometric constraints (e.g., no collisions with walls/doors/windows), and a Refiner adjusts placements based on evaluations. These agents operate iteratively through a Progressive Consensus Mechanism, committing a placement only after passing geometric validation. To support development and evaluation, they introduced InStruct, a benchmark of over 18,000 parametrically annotated interior layouts with structure-specific metrics.

Key findings

• Agentic Designer outperformed state-of-the-art diffusion-based and LLM-based methods on all structure-centric metrics, including collision rate and functional feasibility.
• The Progressive Consensus Mechanism reduced error accumulation, with iterative refinement yielding consistently higher constraint satisfaction than single-pass approaches.
• User studies confirmed that human raters preferred Agentic Designer's layouts for both structural adherence and aesthetic plausibility.

Why it matters

This work demonstrates that decomposing layout generation into modular, verifiable steps can overcome the limitations of black-box one-shot generation. The explicit geometric checking mechanism offers a practical path toward automated spatial design that respects real-world architectural constraints.

Caveats

The framework assumes the availability of parametric room annotations (walls, doors, windows), limiting applicability to settings without such structured input. The iterative refinement process also incurs additional computational cost compared to one-shot methods. The benchmark, while comprehensive, is synthetic; real-world layouts may introduce further complexities not captured in the dataset.