# Function2Scene: 3D Indoor Scene Layout from Functional Specifications

> Function2Scene introduces a novel framework that generates 3D indoor layouts from functional specifications using an iterative check-and-repair pipeline with LLMs, significantly outperforming prior methods in functional design.

- **Source:** [arXiv](https://arxiv.org/abs/2605.30819)
- **Published:** 2026-06-02
- **Permalink:** https://picx.dev/p/hq54Fr
- **Whiteboard:** https://picx.dev/p/hq54Fr/image

## Summary

# Function2Scene: 3D Indoor Scene Layout from Functional Specifications - Summary

## Summary (Overview)
*   **Functionality-First Framing:** Proposes a new paradigm for 3D indoor scene synthesis, shifting the input from object-centric prompts (e.g., "a bedroom with a queen bed") to detailed **functional specifications** (e.g., "a bedroom for a couple where one partner reads late while the other sleeps early").
*   **Constraint Taxonomy:** Introduces a comprehensive, LLM-customizable taxonomy of **17 functional design constraints** organized into four categories: **Spatial (S)**, **Ergonomic (E)**, **Activity (A)**, and **Environmental (N)**, grounded in interior design literature.
*   **Iterative Check-and-Repair Pipeline:** Develops a novel framework that **iteratively evaluates and refines** generated layouts using a tool-augmented loop combining geometric measurements, LLM-based reasoning, and VLM-based visual assessment.
*   **Superior Functional Performance:** Demonstrates that the proposed method significantly outperforms recent LLM-based scene synthesis baselines, with its layouts **preferred in 94.3%** of pairwise comparisons in a perceptual study based on real-world interior design cases.

## Introduction and Theoretical Foundation
A furnished room is not merely a collection of objects but a design that supports human activities. Traditional text-driven 3D scene synthesis methods focus on generating visually plausible object arrangements from prompts that specify *what* furniture to place. In contrast, real interior design starts from a **functional specification**—a natural-language design brief describing *who* will use the space and *what* they need to do there. This work reframes the problem as generating layouts that satisfy these functional requirements.

The paper identifies a gap: while LLMs enable flexible text-conditioned synthesis, they inherit an "implicit" approach, primarily generating scenes from object-centric prompts and checking for visual/physical plausibility, not functional support. The irony is that LLMs are well-suited to the two tasks that limited classical rule-based systems: **parsing open-ended functional descriptions** and **optimizing heterogeneous functional criteria**. This work leverages LLMs to bridge this gap, combining their reasoning power with explicit, customizable design principles.

## Methodology
The **Function2Scene** framework operates in two main stages: **Initialization** and **Constraints-based Evaluation and Refinement**.

### 1. Initialization
Given a raw functional description:
1.  **Parsing:** An LLM parses the input to extract:
    *   A **parsed scene description** (an LLM-friendly reformulation).
    *   A structured list of **functional constraints** customized from the taxonomy based on extracted occupant **personas** and **activities**.
2.  **Room Structure Generation:** An LLM generates an empty room structure (walls, doors, windows) encoded in a custom JSON-based Domain-Specific Language (DSL). This structure is visualized for user verification.
3.  **Furniture Initialization:** An LLM generates an initial furniture layout within the verified room. This serves as a starting point but is often functionally deficient, motivating the refinement stage.

### 2. Constraints-based Evaluation and Refinement
This stage performs an **iterative check-and-repair loop** over the generated layout.

*   **Constraint Taxonomy:** Constraints are organized into four categories across six priority tiers (T1-T6). Lower-tier constraints (e.g., basic spatial validity) must be satisfied before higher-tier ones (e.g., environmental comfort) are considered.

| Category | Constraints | What it checks |
| :--- | :--- | :--- |
| **Spatial (S)** | S1: Geometry Validity<br>S2: Boundary & Attachment<br>S3: Spatial Relationships<br>S4: Scale & Proportion<br>S5: Visual Composition | Object containment, wall attachment, grouping logic, size proportionality, visual balance. |
| **Ergonomic (E)** | E1: Circulation<br>E2: Interaction Clearance<br>E3: Reachability<br>E4: Body Fit & Posture | Clear pathways, door/chair clearance, user reach, anthropometric fit. |
| **Activity (A)** | A1: Activity Zone<br>A2: Sightlines & Privacy<br>A3: Workflow Sequencing<br>A4: Multi-activity Compatibility | Dedicated zones for tasks, unobstructed views, logical activity order, simultaneous activity support. |
| **Environmental (N)** | N1: Natural Light Access<br>N2: Glare Prevention<br>N3: Acoustic Separation<br>N4: Ventilation & Thermal | Daylight access, screen glare avoidance, noise buffering, vent/heat source clearance. |

*   **Evaluation:** For each constraint in priority order, specialized **tools** are invoked to retrieve data, which an LLM interprets to determine satisfaction.
    *   **Geometric/Numeric Tools:** e.g., `boundary_check()`, `pathfinding()`, `free_floor_area()`.
    *   **LLM Query Tools:** e.g., `size_check()`, `reach_check()`, `activity_support_check()`.
    *   **VLM Tools:** e.g., `visual_balance_check()`.
*   **Refinement:** For each unsatisfied constraint, the LLM proposes a targeted refinement action (reposition, reorient, resize) grounded in **design principles** (e.g., minimum 36" circulation path, 2-3' side clearance for a bed).
*   **Termination:** The loop proceeds through all constraints, skipping any that would violate higher-priority, already-satisfied ones. Tier 1 constraints are re-evaluated at the end to ensure foundational quality.

## Empirical Validation / Results
*   **Data:** 30 professionally written interior design cases curated from *Architectural Digest*, spanning 10 room types and 30 unique personas (e.g., retired couple, child with autism, YouTuber).
*   **Baselines:** Compared against three representative LLM-based methods: **Holodeck** [Yang et al. 2024c], **iDesign** [Çelen et al. 2024], and **LayoutVLM** [Sun et al. 2025a].
*   **Perceptual Study:** A two-alternative forced-choice (2AFC) study with 30 participants evaluated which layout better satisfied the functional brief.

**Table 2: 2AFC Study Results Comparing Our Method with Baselines**
| Method | Prompt | % Ours Preferred |
| :--- | :--- | :--- |
| Holodeck [Yang et al. 2024c] | Functional | 92.2 |
| | Parsed | 88.9 |
| iDesign [Çelen et al. 2024] | Functional | 94.4 |
| | Parsed | 98.9 |
| LayoutVLM [Sun et al. 2025a] | Functional | 96.7 |
| | Parsed | 94.4 |
| **Overall** | **—** | **94.3** |

*   **Ablation Study:** Investigated the contribution of pipeline components.

**Table III: 2AFC Study Results Comparing Against Ablations**
| Prompt Format | Iterative Update | Evaluation Tools | % Ours Preferred |
| :--- | :--- | :--- | :--- |
| Functional | No | No | 83.3 |
| Parsed | No | No | 83.3 |
| Functional | Yes | No | 78.9 |
| Parsed | Yes | No | 80.0 |
| **Parsed** | **Yes** | **Yes** | **Ours** |

Key findings:
1.  **Function2Scene is strongly preferred** over all baselines under both original functional and parsed prompts.
2.  **Iterative refinement without evaluation tools is detrimental**, performing worse than no iteration.
3.  **The full pipeline (parsed input + iterative update + evaluation tools) is critical** for superior performance.

## Theoretical and Practical Implications
*   **Theoretical:** The work demonstrates that **explicit, customizable functional constraints** combined with **LLM-driven iterative refinement** can overcome the limitations of purely implicit, data-driven or direct LLM generation approaches. It successfully revisits classical design rule formalization in the era of foundation models.
*   **Practical:** The framework aligns 3D scene synthesis closer to **real interior design workflows**, where the goal is to support human activities. It provides a blueprint for developing more **human-centered AI design assistants**. The constraint taxonomy and tool-augmented evaluation loop offer a reusable structure for incorporating functional reasoning into other spatial AI tasks.

## Conclusion
**Function2Scene** presents a framework for generating 3D indoor layouts from functional specifications. By focusing on functionality, employing a comprehensive and customizable constraint taxonomy, and implementing an LLM-driven iterative check-and-repair pipeline, it produces higher-quality, more functional scenes than prior LLM-based methods.

**Future Directions:**
1.  **Upstream Conversational Interface:** Developing a dialogue system to help non-expert users articulate and refine their needs into detailed functional specifications.
2.  **Enhanced Verification Tools:** Incorporating more powerful tools like embodied simulation with articulated models, physically accurate lighting/acoustic estimation, or a richer DSL for semantic spatial requirements.
3.  **Co-optimization of Architecture:** Extending the framework to jointly optimize room shape, openings, and partitions alongside furniture placement, capturing the full scope of interior design.

---

_Markdown view of https://picx.dev/p/hq54Fr, served by PicX — AI-generated visual whiteboard summaries of research papers._