PaperlayerView exact revision on arXiv

Source-linked AI summary

Function2Scene: 3D Indoor Scene Layout from Functional Specifications

Ruiqi Wang, Qimin Chen, Daniel Ritchie, Angel X. Chang, Manolis Savva, Kai Wang, Hao Zhang

arXiv:2605.30819v1cs.CVcs.GR

TL;DR

Object-centric scene synthesis often specifies furniture without expressing how occupants use a room, although real design briefs center on activities and physical needs. Function2Scene parses those needs into customized constraints and iteratively evaluates and repairs layouts with geometric, language-model, and visual tools. Its layouts were preferred in 94.3% of pairwise comparisons, while the method remains scoped to detailed professional specifications and a verification protocol relying heavily on numeric checks and LLM queries.

  • Problem

    Existing text-driven scene methods largely use object-centric prompts and emphasize visual or physical plausibility rather than functional requirements describing occupants’ activities and needs.

  • Method

    Function2Scene parses personas and activities, derives customized constraints from four taxonomy categories, and iteratively evaluates and repairs layouts using specialized tools.

  • Results

    94.3% aggregate preference was achieved over baselines and ablations in pairwise perceptual comparisons.

  • Takeaways & Limitations

    A functional design taxonomy combined with an LLM-driven iterative pipeline can produce higher-quality, more functional scenes than prior works.

  • Takeaways & Limitations

    The method starts from professionally written, detailed specifications, and its verification protocol heavily relies on basic numeric checks and LLM queries.

Abstract

from arXiv · show

Most text-driven 3D indoor scene synthesis methods generate rooms from object-centric prompts, asking what furniture should be placed rather than how the space is used. Yet in real interior design, a layout is judged by how well it supports its occupants, e.g., their activities and physical needs. We introduce Function2Scene, a framework for generating 3D indoor layouts from functional specifications, i.e., natural-language design briefs describing who will use a room and what they need to do there. Given such a specification, our system parses occupant personas and activities, derives a customized set of functional design constraints from a taxonomy of 17 criteria spanning spatial, ergonomic, activity, and environmental considerations, and uses these constraints to guide layout generation. Rather than relying on an LLM to directly produce a final scene, Function2Scene performs iterative evaluation and refinement through a tool-augmented check-and-repair loop, combining geometric measurements, LLM-based contextual reasoning, and VLM-based visual assessment. Experiments on 30 professionally written interior-design cases show that Function2Scene produces layouts that better satisfy functional requirements than recent LLM-based scene synthesis baselines, with our results preferred in 94.3% of pairwise comparisons. Our work reframes text-driven indoor scene synthesis from placing plausible objects to designing spaces that support human use.

1 Introduction

Function2Scene reframes indoor scene synthesis around functional specifications describing occupants, activities, and needs rather than merely requested objects. It parses these needs into constraints and iteratively refines layouts, achieving a 94.3% aggregate pairwise preference rate.

  • Object-centric LLM scene methods mainly optimize visual quality and physical plausibility, leaving functionality-oriented prompts underexplored.
  • Functional specifications describe who uses a space, what they do, and which constraints their needs impose.
  • Function2Scene customizes design guidelines from 17 criteria across Spatial, Ergonomic, Activity, and Environmental categories using occupant personas and activities.
  • 94.3% aggregate pairwise preference was achieved over baselines and ablations in a crowdsourced 2AFC perceptual study.
  • The framework combines geometric measurements, LLM reasoning, and VLM visual assessment in a tool-augmented check-and-repair loop.

2 Related Works

Prior indoor scene synthesis progressed from manually authored rules and spatial priors to flexible LLM-based and iterative systems. Despite these advances, most methods remain object- and spatial-plausibility-oriented rather than explicitly functionality-oriented.

  • Pre-LLM methods relied on prespecified principles, statistical relationships, and hand-written programs requiring substantial manual effort.
  • LLM-based systems expanded text-conditioned scene generation from object coordinates to embodied environments, personalization, and richer spatial reasoning.
  • Across this progression, prompts predominantly specify objects, relations, and coordinates while optimization emphasizes spatial plausibility and visual coherence.
  • Recent agentic and VLM-guided pipelines iteratively refine layouts after initial generation using feedback or editing.

3 Design Constraints

Function2Scene represents interior design requirements as persona- and activity-dependent constraints organized across spatial, ergonomic, activity, and environmental categories. These constraints are prioritized and evaluated with specialized tools to capture how layouts support use.

  • Activity × persona combinations determine which constraints apply and how their thresholds are parameterized for a scene.
  • The taxonomy organizes constraints into Spatial, Ergonomic, Activity, and Environmental categories grounded in interior-design literature.
  • Ergonomic constraints address circulation, interaction clearance, reachability, and body fit according to users’ physical needs and abilities.
  • Activity constraints allocate usable zones, support multi-activity transformation, encode sightlines and privacy, and order objects by workflow.
  • Environmental constraints address daylight, glare, acoustic separation, and ventilation or thermal comfort.
  • Constraints are assigned six priority tiers, with lower tiers satisfied before higher tiers are considered.

4 Method

Function2Scene converts a functional prompt into a structured room and initial furniture layout, then evaluates and repairs constraint violations in priority order. The method treats initialization as a starting point because direct LLM layouts can remain practically unusable.

  • The function-driven agent produces furniture layouts tailored to occupants’ living needs and use of the space.
  • Initialization: Initialization parses the prompt, generates a verifiable room structure, and creates an initial furniture layout.
  • Initialization: Initial LLM-generated layouts may overlap, violate functional adjacency, or be physically plausible yet practically unusable.
  • Constraints-based Evaluation and Refinement: The evaluation loop invokes specialized tools and lets the LLM interpret numeric, spatial, and visual feedback for each constraint.
  • Constraints-based Evaluation and Refinement: For unsatisfied constraints, the LLM generates targeted actions grounded in universal spatial standards and room-specific human-factor recommendations.
  • Constraints-based Evaluation and Refinement: Sequential termination rechecks Tier 1 spatial constraints after later adjustments to ensure foundational layout quality is preserved.

5 Results and Evaluation

The evaluation compares Function2Scene with LLM-based baselines and ablations on professionally curated interior-design cases using a 2AFC perceptual study. Participants preferred the proposed method across baselines, while ablations show that grounded evaluation tools are critical for effective iterative refinement.

  • Evaluation Setup: The study uses 30 Architectural Digest interior-design cases spanning diverse room types and occupant personas.The cases were curated as professionally written functional specifications.
  • Evaluation Setup: Participants compared rendered layouts in a randomized two-alternative forced-choice study across baseline and ablation conditions.The evaluation included 30 participants, each completing 30 scene comparisons.
  • Perceptual Results: 94.3% aggregate preference favored Function2Scene across all baselines and prompt conditions.Preference rates were 92.2% and 88.9% against Holodeck, 94.4% and 98.9% against iDesign, and 96.7% and 94.4% against LayoutVLM under functional and parsed prompts.
  • Perceptual Results: Function2Scene was preferred in every baseline comparison, including Holodeck, iDesign, and LayoutVLM.The reported rates vary by baseline and prompt condition, with parsed prompts producing the highest comparison score against iDesign.
  • Ablation Results: Iterative updates performed worst without evaluation tools, showing that refinement requires grounded spatial feedback.Prompt format had negligible effect when tools were absent, while richer constraints helped when paired with tools that measured them.

6 Conclusion

Function2Scene frames indoor layout generation around functional specifications and combines a functional-principle taxonomy with an iterative LLM-driven pipeline. The authors report higher-quality, more functional scenes than prior works while identifying broader workflow and verification extensions.

  • Conclusion: Function2Scene generates indoor layouts from functional specifications rather than solely object-centric descriptions.The framework is presented as a step toward LLM-driven scene generation suited to real interior-design workflows.
  • Conclusion: A taxonomy of functional design principles combined with an LLM-driven iterative pipeline produces higher-quality, more functional scenes than prior works.This is the paper’s stated conclusion about the combined design.
  • Limitations and Extensions: The current method starts from professionally written, detailed specifications, whereas real workflows often begin with vaguer user demands.The authors identify conversational need discovery and refinement as an upstream extension.
  • Limitations and Extensions: The verification protocol relies heavily on basic numeric checks and LLM queries within fixed residential architectural shells.The authors suggest constraint-specific tools and broader architectural settings as extension opportunities.

occasionally

Figure 5 presents functional scenes generated by the method with zoomed-in highlights, while supplementary materials contain fuller qualitative results and intermediary optimization visualizations.

  • Figure 5: Figure 5 shows generated functional scenes alongside zoomed-in highlights.The input prompts are truncated in the figure.
  • Figure 5: The supplementary materials provide all qualitative results and visualizations of intermediary optimization steps.The figure caption directs readers to the supplementary materials for the complete visual record.

A Layout Representation

Function2Scene represents room layouts with a JSON-based domain-specific language designed for human authoring, automated validation, and 3D rendering. The representation separates architectural structure from movable furniture and records shared geometric conventions and orientation semantics.

  • DSL Overview: The layout DSL is encoded in JSON as a structured, machine-readable representation of room architecture and furnishings.It is designed for both human authoring and automated layout validation and rendering.
  • DSL Structure: The DSL separates architectural surfaces in room_structure from movable objects in furniture.Architectural surfaces include walls, floor, ceiling, door, and window.
  • Geometric Conventions: Each surface and object stores a world-space centroid and local dimensions expressed as width, height, and depth in metres.The representation uses a shared coordinate convention defined in each file’s convention block.
  • Orientation Semantics: Furniture orientation tags distinguish directional, axial, and symmetric objects for rendering and constraint validation.These categories encode whether an object has a meaningful front, one-axis symmetry, or rotational symmetry.

B Perceptual Study Details

The perceptual study compared Function2Scene layouts with baseline methods and ablations across 30 room scenes using pairwise participant judgments. Participants viewed room briefs alongside rendered scene pairs and evaluated layout quality across four dimensions.

  • Study Design: 30 room scenes were evaluated in a two-alternative forced-choice study against outputs from 10 baseline methods or ablations.Each survey contained 30 pairs, with each pair drawn from a different scene.
  • Study Design: 5 of 30 pairs were attention checks against randomly generated layouts, and participants failing any check were excluded.The checks were distributed throughout each survey.
  • Study Design: 10 survey variants assigned comparison methods across scenes, with 3 participants targeted per variant.The variants were presented through a web-based interface showing a room brief and two rendered layouts.
  • Participants: 45 submissions were collected, with 32 passing all attention checks and retained for analysis.The valid response rate was 71.1%, and two repeated results were removed.
  • Interface: Participants selected the better layout after viewing the room brief, persona description, and two rendered scenes side by side.The interface also provided reminders for spatial validity, ergonomics, activity support, and environmental quality.

C.1 Initialization

The initialization pipeline converts a natural-language room description into a structured scene through staged LLM transformations and human verification. It separately generates the room shell, openings, and furniture while preserving structural consistency.

  • Initialization: The pipeline transforms natural-language room descriptions into structured JSON through discrete, narrowly scoped LLM-driven steps.Intermediate outputs are saved so a human reviewer can inspect and correct them before the next stage.
  • Room Shell Generation: Room shell generation classifies descriptions as rectangular or L-shaped and generates the corresponding structural envelope.Structural cues such as alcoves, nooks, missing corners, or right-angle zones trigger the L-shaped path.
  • Room Shell Generation: The shell uses named walls, a floor polygon or slab, and a ceiling, while leaving all wall-opening arrays empty initially.No openings are cut into walls during this stage.
  • Door and Window Placement: Door and window placement appends openings to the room structure and records matching wall holes in local wall coordinates.Placement constraints ensure openings fit within wall extents and account for wall orientation.
  • Human Verification: Human verification pauses the pipeline after opening placement so reviewers can correct partitions, openings, or dimensions before furniture generation.This checkpoint prevents furniture placement against an incorrect structural baseline.
  • Furniture Generation: Furniture generation reads the reviewed structure and adds object entries without modifying structural elements.Zone checklists ensure essential objects populate implied sleeping, living, kitchen, bathroom, work, and dining areas.

C.2 Functional Constraints

The functional-constraint system parses a functional prompt into a scene description and customized constraints covering spatial validity, attachment, circulation, access, orientation, and environmental interactions. Constraints specify pass conditions and targeted repairs for violations.

  • Constraint Parsing: The parser generates both a parsed scene description and functional constraints from the functional prompt.The constraint specification is organized into zone maps, rule groups, priorities, pass conditions, and repair actions.
  • Tier 1 - Hard: Spatial rules prevent object overlap and boundary violations while enforcing floor contact and valid wall attachment.Exceptions allow only explicitly defined containment or paired-object relationships.
  • Tier 1 - Hard: Wall-mounted objects must lie entirely on solid wall segments and avoid windows, doors, and corner voids, with lateral relocation or replacement when necessary.Curtain rails are the stated exception to the opening-clearance rule.
  • Tier 1 - Hard: Door rules require clear swing arcs, unobstructed approach paths, and passable open thresholds.Repairs move blocking objects or widen narrow threshold paths by shifting flanking objects outward.
  • Tier 1 - Hard: Window rules prevent non-curtain objects from blocking glazing and invoke visual assessment for low units near sill lines.Curtains and curtain rails are exempt, while radiators or built-ins below sills may require manual review.
  • Tier 2 - High: Layout Plausibility: Ergonomic and activity rules require accessible fixture orientation, passable circulation, and reachable sleeping surfaces for intended users.Fixture use-faces must point toward unobstructed floor, while deployed sleeping surfaces require access along both long sides.

C.3 Evaluation Tools

Function2Scene evaluates layouts with geometric, path-based, object-state, contextual, and visual tools. These checks measure validity and clearances while using LLM judgments for human factors, activities, workflows, environmental risks, and visual balance.

  • Geometric Checks: The evaluation toolkit includes geometric checks for boundaries, wall angles, object existence, object state, size ratios, and bounding-box collisions.These tools return Boolean, numeric, or structured object-state results for scene validation.
  • Visual Assessment: Visual-balance assessment uses a VLM to judge a top-down rendered view for balance, focal point, and alignment.The result is a free-text assessment rather than a numeric score.
  • Navigation and Clearance: Pathfinding and clearance tools compute navigable routes, minimum path widths, door or drawer articulation clearance, and chair clearances.Pathfinding uses A* over floor positions and reports waypoints or failure when no path is found.
  • Activity Checks: Multi-activity checks assess whether a scene supports activities simultaneously and identify incompatible activity pairs.The check reuses free-floor, zone-placement, and activity-support evaluations.
  • Environmental Checks: Environmental tools evaluate window and vent obstruction, screen-window geometry, glare risk, zone distance, and acoustic separation.These tools combine geometric measurements with LLM judgments of glare and acoustic risk.
Loading 2605.30819v1…