Function2Scene 3D Indoor Scene Layout from Functional Specifications

The figures below are drawn from Julius Panero and Martin Zelnik's foundational reference work, Human Dimension and Interior Space: A Source Book of Design Reference Standards (Whitney Library of Design / Watson-Guptill Publications, New York, 1979). The book applies anthropometric data—measurements of the human body in motion and at rest—to the practical challenges of interior space planning. These four diagrams focus on residential and dining environments: the first two address lounge seating, illustrating the dimensional requirements for sofas accommodating two and three occupants, and the clearance zones needed around corner lounge arrangements to allow comfortable circulation. The latter two turn to dining and kitchen spaces, specifying the minimum clearances around a breakfast table for four, and the sitting, circulation, and standing activity zones required between a dining table and a buffet or wall. Together, they exemplify the book's core argument: that well-designed interior space is not a matter of taste alone, but of precise, human-centered measurement.

In practice, layout requirements are not universal. A mobility-limited user and a frequent entertainer sharing the same room type will impose fundamentally different demands on furniture placement, clearance, and zone allocation. We capture this variability through activity × persona combinations derived from the user's functional specifications, which together determine which constraints are relevant and how their thresholds are parameterized for a given scene.

We ground our constraints taxonomy in established interior design literature and organize constraints into four categories: Spatial, Ergonomic, Activity, and Environmental.

To verify each constraint at runtime, our system employs 29 specialized tools across three types. Numeric and geometric tools compute quantitative measures directly from scene geometry — overlap ratios, traversal paths, clearance distances, and obstruction percentages. LLM query tools apply language model reasoning over structured scene data for contextual judgments that resist straightforward numeric formulation, such as workflow order, activity support, and glare risk. VLM tools interpret rendered images for holistic visual assessments, used here for top-down balance evaluation. Constraints are organized into six priority tiers (T1–T6): lower-tier constraints must be satisfied before higher-tier ones are evaluated, ensuring that geometric validity is confirmed before ergonomic fit, and ergonomic fit before activity-level behavior.

Given a functional prompt capturing the user's living needs, the system proceeds through two main stages. In the Initialization stage, the prompt is parsed into a structured constraint list and a reformulated scene description; the room structure is then generated and verified by the user before an initial furniture layout is produced. Because LLM-generated layouts at this stage frequently suffer from overlaps, violated adjacencies, and spatially plausible but practically unusable configurations, they serve as a starting point rather than a final result. The second stage, Constraints-based Evaluation and Refinement, addresses this by iteratively assessing the layout against all 17 constraints in priority order — invoking the specialized tools described above, generating a targeted adjustment for each unsatisfied constraint, and re-evaluating before moving on — until the layout satisfies functional, ergonomic, activity, and environmental requirements end-to-end.

We evaluate Function2Scene through a two-alternative forced-choice (2AFC) perceptual study with 30 participants recruited via Prolific, comparing against Holodeck, iDesign, and LayoutVLM across 30 real interior design cases spanning 10 room types and 30 distinct occupant personas sourced from Architectural Digest. Participants were shown two rendered layouts side by side and asked to select the more functional one for the described occupant. Across all baselines and prompt conditions, participants preferred Function2Scene at an aggregate rate of 94.3%, with the margin reaching as high as 98.9% against iDesign under our parsed prompt. An ablation study further identifies the evaluation tool set as the critical enabler: applying iterative refinement without tools performs worse than applying no refinement at all, confirming that grounded spatial feedback, not iteration alone, drives the improvement.

We invite readers to explore the full set of 30 generated scenes in our interactive viewer, where each layout can be inspected from any angle alongside its parsed prompt and constraint list.