Skip to content

Co-Layout: LLM-driven Co-optimization for Interior Layout

🕒 Published (v1): 2025-11-16 06:20 UTC · Source: Arxiv · Venue: AAAI 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Co-Layout combines a multi-agent LLM workflow with grid-based integer programming to jointly optimize room layout and furniture placement from a text prompt. Unlike prior two-stage pipelines, it co-optimizes both levels simultaneously, enforcing hard spatial constraints (corridor connectivity, room accessibility, furniture containment) via an IP solver. It outperforms Holodeck and AnyHome on all measured metrics including zero object overlap and boundary violations.

Problem

Existing automated interior design systems decouple room layout generation from furniture placement into sequential stages, preventing synergistic optimization—furniture affects room sizing, and rooms constrain furniture. LLMs alone cannot produce precise coordinates and frequently generate conflicting placements. Neither rasterized nor vectorized floorplan representations can natively encode high-level topological constraints like corridor connectivity or room accessibility.

Method

A multi-agent LLM preprocessing pipeline (specialized agents for basic info extraction, environment analysis, room configuration, and furniture specification) converts a freeform text prompt into a structured JSON scene graph of spatial constraints. This scene graph is encoded into a unified grid-based representation (cells carry binary indicator variables for room x^k_{i,j}, corridor p_{i,j}, and furniture f^{k,l}_{i,j} assignment) inspired by Le Corbusier's "Modulor" system. The resulting mixed-integer program (solved with Gurobi) co-optimizes room layout and furniture placement subject to: non-overlapping (mutual exclusivity over cells), corridor connectivity (flow-based formulation ensuring a single connected corridor network), room accessibility (every room must have a furniture-free access point adjacent to the corridor), room adjacency (shared boundary enforcement), and furniture containment, shape, and non-overlap constraints. Objective terms penalize bounding-box area deviation (rectangularity), exposed perimeter, area target deviation, aspect ratio extremes, furniture relative positioning error, and centroid imbalance. A coarse-to-fine strategy first solves room layout on a spatially downsampled grid (e.g., 6Ă—5 from 12Ă—10), maps the solution back as a warm start, then runs the full IP with furniture variables at fine resolution, adding soft penalty terms for consistency with the coarse assignment.

Key Contributions

  • Joint co-optimization of room layout and furniture placement in a single IP, replacing sequential two-stage pipelines.
  • Grid-based representation with formally specified constraints for corridor connectivity (flow conservation), room accessibility, and inter-room adjacency, all LLM-derived.
  • Coarse-to-fine optimization strategy that provides warm-start initialization and reduces runtime significantly as grid resolution grows.

Results

  • Object Overlap Rate (OOR): Ours 0.00 vs. AnyHome 0.00 vs. Holodeck 0.82 ↓ (lower is better)
  • Out-of-Boundary Rate (OOB): Ours 0.00 vs. AnyHome 0.04 vs. Holodeck 2.33 ↓
  • Image Quality (IQA, Q-Align): Ours 4.17 vs. AnyHome 4.10 vs. Holodeck 4.03 ↑
  • Image Aesthetics (IAA, Q-Align): Ours 3.35 vs. AnyHome 3.32 vs. Holodeck 3.32 ↑
  • CLIP text-image alignment: Ours 26.50 vs. AnyHome 25.75 vs. Holodeck 25.15 ↑
  • User study (n=64): Ours scores highest on semantic alignment (3.77), layout rationality (3.23), path clearance (3.41), and MRR (0.80 vs. 0.59/0.45 for Holodeck/AnyHome)
  • Coarse-to-fine reduces solve time substantially at larger grid sizes (exact speedup not quantified numerically, shown via ablation chart)

Limitations

  • Furniture scope restricted to floor-placed objects; wall-mounted and tabletop objects are unsupported.
  • LLM agents may produce internally conflicting constraints, degrading IP solution quality; no automatic conflict detection or resolution.
  • Framework does not generalize to multi-story buildings; inter-floor constraints are absent.
  • Evaluation conducted on only five layout examples for quantitative metrics; limited scale.

Relevance to Agentic AI / LLM Agents

Co-Layout is a concrete instance of the hybrid agentic pattern where LLMs act as structured specification generators rather than end-to-end solvers: a multi-agent pipeline decomposes a complex design task into specialized semantic analysis subtasks, then hands structured output to a formal solver. This illustrates how LLM agents can serve as the natural-language-to-constraint interface in tool-augmented agent architectures, offloading combinatorial feasibility guarantees to a downstream optimizer. The multi-agent decomposition (basic info → environment → rooms → furniture) mirrors hierarchical task decomposition strategies seen in broader agentic AI research. The paper also highlights a key limitation of pure LLM planning—inability to produce globally consistent coordinate assignments—and proposes constraint extraction + solver as a principled remedy.