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

TL;DR

Co-Layout presents a framework that merges LLM-generated design constraints with a grid-based integer programming formulation to jointly optimize interior room layouts and furniture placement. The method encodes high-level user requirements into a Modulor-inspired spatial grid and solves the resulting mixed-integer program with a coarse-to-fine strategy, yielding layouts that satisfy corridor connectivity and room accessibility while optimizing geometric and functional quality. An LLM-based preprocessor converts prompts into structured scene graphs, and a post-processing stage adds walls, doors, and windows, followed by a Blender-based rendering pipeline for visualization. Quantitative and user studies show superior physical plausibility, aesthetics, and text–image alignment compared with two-stage baselines (Holodeck and AnyHome), with substantial computational gains from the coarse-to-fine approach. The work highlights practical interior-design automation and suggests future expansions to wall-mounted furniture, iterative refinement, and multi-story design.

Abstract

We present a novel framework for automated interior design that combines large language models (LLMs) with grid-based integer programming to jointly optimize room layout and furniture placement. Given a textual prompt, the LLM-driven agent workflow extracts structured design constraints related to room configurations and furniture arrangements. These constraints are encoded into a unified grid-based representation inspired by ``Modulor". Our formulation accounts for key design requirements, including corridor connectivity, room accessibility, spatial exclusivity, and user-specified preferences. To improve computational efficiency, we adopt a coarse-to-fine optimization strategy that begins with a low-resolution grid to solve a simplified problem and guides the solution at the full resolution. Experimental results across diverse scenarios demonstrate that our joint optimization approach significantly outperforms existing two-stage design pipelines in solution quality, and achieves notable computational efficiency through the coarse-to-fine strategy.

Figures (8)

• Figure 1: Overview of our automated framework. (a) First, the LLM-based workflow processes user requirements to generate spatial constraints for rooms and furniture, along with boundary conditions for the floor. (b) These constraints are then encoded into an integer programming model using our grid-based representation. The model is efficiently solved through a coarse-to-fine strategy to obtain the final layout. (c) The generated layout is converted into a scene white box in Blender blender. Suitable assets are then retrieved from the 3D-FUTURE fu20213d dataset and a curated library of over 2,000 assets via semantic embedding and size matching to populate the scene.
• Figure 2: The floor is modeled as a 2D grid. Each cell $(i, j)$ has binary variables indicating if it belongs to a corridor ($p_{i, j}$), room $r_k$ ($x_{i, j}^k$), or is occupied by the l-th furniture in the k-th room $s_{k, l}$ ($f_{i, j}^{k, l}$). Each furniture $s_{k, l}$ has dimensions $W_{k, l} \times L_{k, l}$. Its orientation $\boldsymbol{\nu}_{k, l}$, chosen from 4 axis-aligned directions, is set by two binary variables $\sigma_{k, l}$ and $\mu_{k, l}$.
• Figure 3: Various examples generated by our method, demonstrating its capability to handle diverse inputs including residential and non-residential spaces. Users can specify requirements such as building type, floor area, room functions, and required furniture. Top: Input text. Middle: Bird's-eye view of the scene. Bottom: Zoom-in views highlighting key details.
• Figure 4: Comparison with baselines. Our approach consistently generates well-structured layouts with full accessibility and clear circulation. In contrast, baseline methods often produce designs with critical flaws, such as illogical circulation paths that violate privacy (e.g., Holodeck in a, d) and unreachable spaces or impractical room shapes (e.g., AnyHome in c, e).
• Figure 5: Ablation result of the coarse-to-fine strategy.