PaperlayerView exact revision on arXiv

Source-linked AI summary

Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

Fedor Rodionov, Aleksandar Cvejic, Michael Birsak, John Femiani, Peter Wonka

arXiv:2606.10953v2cs.AIcs.CV

TL;DR

Automatic furnishing must address limited real-world data and interacting geometric and functional constraints. Architect-Ant combines professional floor-plan examples, pretrained knowledge, and rule-based outcome rewards to generate editable layouts. It achieves high functional and layout-rule scores with low geometric violation rates, and the layouts can be converted into 3D scenes.

  • Problem

    Automatic furnishing has limited real-world data and must satisfy interacting geometric and functional constraints.

    Source: §1 · pp. 1–2

  • Method

    Architect-Ant learns from AntPlan using supervised fine-tuning, then optimizes editable coordinate-based layouts with GRPO rewards from the Layout Rule Score.

    Source: §1 · p. 2

  • Results

    Architect-Ant achieves the highest mean functional-completeness (94%) and LRS (8.80) among compared methods, with collision and out-of-bounds rates of 0.6% each.

    Source: §5 · p. 9

  • Takeaways & Limitations

    Generated layouts remain object-level editable and can be evaluated geometrically and converted into furnished 3D scenes.

    Source: §5 · p. 9

    LRS measures only geometric and functional properties encoded by its rules; it does not measure aesthetics or guarantee full 3D usability.

    Source: §3.4 · p. 5

Abstract

from arXiv · show

Furnished floor plans support real-estate visualization, interior design, and architectural workflows, yet automatic furnishing remains challenged by limited real-world data and the need to satisfy interacting geometric and functional constraints. We ask whether professional furnishing knowledge can be learned from real floor plans using a pretrained model, enabling direct constraint-aware layout generation without relying on costly iterative agentic inference. We introduce AntPlan, a curated dataset of 505 real professional architectural floor plans with dense furniture annotations spanning 92 object classes and ten residential room categories, and Architect-Ant, a framework for generating furniture layouts. Architect-Ant represents layouts with an editable coordinate-based DSL and first learns professional furnishing patterns through supervised fine-tuning. It is then optimized with GRPO using a Layout Rule Score (LRS) that aggregates geometric and functional constraints derived from professional plans, providing outcome-level supervision without prescribed reasoning traces. Experiments against diverse state-of-the-art baselines show that Architect-Ant combines low geometric violation rates with high functional completeness, while qualitative results more closely reflect real-world residential furnishing patterns. The resulting layouts remain object-level editable and can be converted into 3D scenes.

1 Introduction

The introduction frames automatic furnishing as a problem of satisfying interacting geometric and functional constraints, then presents a dataset and a framework for learning constraint-aware layouts. Architect-Ant achieves high functional completeness and overall layout-rule scores while maintaining low geometric violation rates and generating layouts faster than iterative agentic approaches.

1 Introduction

  • Furniture placement must satisfy interacting geometric and functional constraints, such as fitting objects while preserving access, usable arrangements, and circulation.Moving one object can improve one relationship while breaking another, and many requirements are implicit in professional practice.Source: §1 · p. 1
  • The framework combines layout examples, explicit spatial constraints, and pretrained-model knowledge to learn constraint-aware generation during training.The introduction contrasts this direction with iterative agentic refinement, whose rounds increase inference cost and time.Source: §1 · p. 2
  • AntPlan contains 505 real professional floor plans with dense furniture annotations across 92 object classes and ten residential room categories.The dataset addresses limitations of projected 3D layouts, which may not reflect professional floor-plan conventions and construction elements.Source: §1 · p. 2
  • Architect-Ant uses an editable coordinate-based DSL and combines supervised adaptation on professional layouts with GRPO rewards from geometric and functional constraints.Rewards are applied to final layouts rather than prescribed reasoning traces.Source: §1 · p. 2
  • Across four room types, Architect-Ant achieves the highest functional-completeness and overall layout-rule scores while maintaining low geometric violation rates and generating layouts substantially faster than iterative agentic approaches.The introduction reports the overall comparison without giving metric values.Source: §1 · p. 2

2 Related Work

Related work spans architectural floor-plan generation, furnishing and scene synthesis, language-guided planning, and rule-guided learning. Architect-Ant differs through its use of professional floor-plan annotations and spatial feedback during training.

2 Related Work

  • Prior floor-plan datasets and indoor-scene datasets provide structural, room, or object annotations, whereas AntPlan targets dense, fine-grained furniture labels on real professional drawings.Examples include CubiCasa5K, FloorPlanCAD, 3D-FRONT, and 3D-FUTURE (Fu et al., 2020a;b).Source: §2 · p. 3
  • Furniture arrangement methods range from classical geometric and ergonomic optimization to learned models trained on furnished 3D-FRONT scenes, while Architect-Ant learns from real architectural drawings.Examples of learned approaches include ATISS, DiffuScene (Tang et al., 2023), and InstructScene.Source: §2 · p. 3
  • Language-guided scene-planning systems use retrieval, spatial optimization, or hierarchical agents; Architect-Ant instead trains with spatial feedback and generates candidates without a model-based critique-and-revision loop.The section also relates its scalar geometric and functional rewards to verifier-based learning and GRPO.Source: §2 · p. 3

3 Architect-Ant

Architect-Ant combines pretrained model knowledge, professional floor-plan examples, and spatial constraints to generate editable coordinate-based furniture layouts and convert them into 3D scenes.

3 Architect-Ant

  • The system combines pretrained knowledge, AntPlan supervision, and spatial constraints to generate coordinate-based layouts, then converts selected layouts into 3D scenes.Room geometry, a structure-only image, and a furniture request condition generation.Source: §3 · p. 3, Figure 4 · p. 6, §3.6 · p. 6

3.1 AntPlan and room inputs

  • AntPlan supplies 92 furniture classes averaging 53.01 annotated objects per plan; experiments use four room types and source-plan-exclusive training and validation splits.Structural elements are detector-extracted, room labels are assigned manually, and furniture annotations are manually reviewed and corrected.Source: §3.1 · p. 4

3.2 Editable layout representation

  • The editable DSL records each object’s class, top-left footprint corner, and dimensions in meters relative to the room’s top-left corner.Footprints support direct containment, overlap, and clearance checks; downstream conversion infers orientation from furniture relationships and wall attachment.Source: §3.2 · p. 4

3.3 Supervised adaptation

  • Supervised fine-tuning adapts a pretrained model to furnishing and the DSL using final-layout targets, without reasoning traces.Training uses token-level cross-entropy; the vision encoder remains frozen while the language model is fine-tuned with LoRA.Source: §3.3 · p. 4

3.4 Layout Rule Score and reward

  • LRS starts each layout at 10 and deducts penalties across 33 shared and room-specific rule types, with thresholds calibrated against AntPlan annotations.Training adds auxiliary rewards, but inference ranks candidates by raw LRS; the score covers only encoded geometric and functional properties, not aesthetics or full 3D usability.Source: §3.4 · pp. 4–5

3.5 GRPO with reasoning from layout rewards

  • GRPO rewards parsed final layouts only, while applying room-group outcome advantages to reasoning and layout tokens without prescribing the reasoning process.Each room contributes 16 responses; batches contain eight rooms and receive one policy-gradient step, with no learned critic.Source: §3.5 · p. 5

3.6 Candidate selection and 3D rendering

  • For a new room, the system retrieves dimensionally similar same-type rooms to form furniture requests, generates layouts, and selects the highest raw-LRS candidate.The selected DSL remains editable; compatible 3D assets are fitted to predicted footprints, with orientation inferred from walls and furniture relationships.Source: §3.6 · p. 6

4 Experiments

Experiments evaluate training variants on held-out AntPlan rooms and compare furnishing pipelines on shared SceneSmith shells. Architect-Ant leads in functional completeness and overall layout-rule scores while maintaining low geometric violation rates.

4.1 Benchmark and comparison protocol

  • The SceneSmith comparison uses 109 shared room shells; Architect-Ant receives four furniture requests per room, while methods use distinct generation and refinement pipelines.SceneSmith’s released scenes include native iterative refinement, and the compared pipelines differ in candidate counts and internal search.Source: §4.1 · p. 6

4.2 Metrics and comparison results

  • Architect-Ant achieves the highest mean FUN+ and LRS, low collision and out-of-bounds rates, and the highest mean furniture count per room.Its accessible-pair rate is 97.2%, within 2.3 points of Holodeck’s 99.5%.Source: Table 3 · p. 7, §4.2 · p. 7
  • Coverage measures furnishing occupancy rather than quality, so geometric and functional usability requires jointly considering collision, boundaries, access, completeness, and LRS.LRS also includes relationships not captured by collision or containment alone.Source: §4.2 · p. 7

4.3 VLM-based visual preference

  • Figure 5 gives qualitative 3D comparisons across four room types, while Table 4 reports pairwise preferences as wins, ties, and losses plus mean LRS.Learned 3D-FRONT baselines appear only for room categories they support.Source: Figure 5 · p. 8, Table 4 · p. 9

4.4 Effect of SFT and GRPO

  • On 243 held-out AntPlan rooms, reasoning-enabled GRPO after SFT improves both mean and best-of-eight LRS across all four room types.The comparison includes the base model, direct DSL output, five-epoch direct-DSL SFT, and SFT followed by GRPO.Source: Table 5 · p. 9, §4.4 · p. 9

5 Conclusion

Architect-Ant combines supervised learning from AntPlan with LRS-rewarded GRPO to generate editable layouts. It achieves strong geometric and functional results, with further extensions identified for future work.

5 Conclusion

  • Supervised fine-tuning on AntPlan teaches furnishing patterns, and GRPO improves layout quality using LRS rewards without supervised reasoning traces.The layouts remain editable and can be evaluated geometrically and converted into furnished 3D scenes.Source: §5 · p. 9
  • Architect-Ant achieves the highest mean functional-completeness (94%) and LRS (8.80), with collision and out-of-bounds rates of 0.6% each.Both visual judges prefer its layouts over SceneSmith in 72 and 74 of 109 paired rooms, respectively.Source: §5 · p. 9
  • Future work can extend the framework to additional object and scene properties, richer geometry, and more detailed spatial relationships.The conclusion calls for preserving structured, editable generation as these extensions are pursued.Source: §5 · p. 9

A AntPlan dataset

AntPlan links professional floor-plan images with structural, room, and object annotations, creating room-level furnishing examples for the experiments.

A AntPlan dataset

  • AntPlan pairs original architectural drawings with structural boxes, object labels, and room regions across ten room categories and 92 object classes.Figure 6 illustrates the original plan, labeled structural and object boxes, and categorized room regions.Source: A AntPlan dataset · p. 13, Figure 6 · p. 13

A.1 Annotation workflow

  • Annotation combines detector proposals for structure and furniture with manually assigned room labels and manual review of furniture detections.The reviewed furniture annotations are treated as pseudo-labels because symbols can be missed, classes can be ambiguous, and box boundaries imperfect.Source: §A.1 · p. 13

A.2 Dataset statistics

  • The cleaned snapshot contains 505 plans, 4,495 room annotations, and 26,716 object annotations, including 715 objects not assigned to a room.The snapshot averages 8.90 rooms per room-labeled plan and 53.01 objects per object-labeled plan.Source: §A.2 · p. 13

A.3 Saved records and room extraction

  • Room crops use a 0.5 m context margin and approximate metric scale from median door width, assuming a 0.8 m opening; plans without usable door-based scale are excluded.Derived room records retain category, scale estimate, geometry, and furniture boxes linked to structure-only images and target DSL layouts.Source: §A.3 · p. 14

A.4 Experimental splits and class whitelists

  • Furnishing experiments filter four room types to 53 distinct DSL classes, require at least one retained bathroom object and two in other rooms, and split examples by source plan within each room type.These experimental whitelists are narrower than the full annotation vocabulary; a source plan may fall in training for one room adapter and validation for another.Source: §A.4 · p. 14, Table 7 · p. 15, Table 8 · p. 15

B Training and reward implementation

Room-specific supervised fine-tuning teaches the model to emit furniture DSL, followed by group-relative optimization using a reward based on rule scores and reference-layout similarity. The training setup also specifies room-level data splits, reward handling for unparseable layouts, and token-masked loss accumulation.

Data and supervised training

  • The cleaned data contain 2,259/243 training/validation room examples overall, with parent-floorplan-disjoint splits within each room type.Source plans may cross partitions between separately trained room types.Source: B Training and reward implementation · p. 15
  • Separate room-type language-model LoRA adapters are trained with the vision encoder frozen, using DSL cross-entropy while masking prompt, image, and padding tokens.Training uses no geometric augmentation; each example includes a structure-only raster, textual geometry, requested inventory, and annotated furniture DSL.Source: B Training and reward implementation · p. 15

Training reward

  • Unparseable furniture receives −40; otherwise the optimization reward is q + 2σ((q − 7)/1.5)G, where q = S − 2B.The optimization reward is distinct from the reported LRS score.Source: B Training and reward implementation · p. 15
  • For G, reference objects are greedily matched to nearest unused predicted centers of the same class, and unmatched references contribute zero.For B, objects are paired within each class by increasing area, with a pair counted when f < 0.5; inventory penalties handle missing and surplus objects.Source: B Training and reward implementation · p. 16

Group-relative optimization

  • Each iteration samples fresh reasoning-plus-layout completions and makes one optimizer update; accumulated loss uses token-budget microbatches and completion-token masks.Groups with identical rewards are skipped, and the implementation is group-relative policy gradient rather than a PPO-clipped importance-ratio surrogate.Source: B Training and reward implementation · p. 16

C Layout Rule Score

The Layout Rule Score aggregates geometric and functional constraints inferred from annotated professional plans into a proxy for plausible furniture arrangements. Its checks use room geometry and the proposed layout, with room-specific rules and tolerances.

Score design

  • LRS aggregates explicit geometric and functional constraints inferred from professional-plan annotations and can be evaluated without a reference furniture arrangement.The separate per-example ground-truth similarity term is added to the training reward, not to the LRS itself.Source: C Layout Rule Score · pp. 16–17

Penalty design

  • The principal penalty families deduct costs from 10, and multiple violations can accumulate.Small annotation tolerances are intended to suppress extraction noise, not to represent building-code requirements.Source: Table 10 · p. 17, C Layout Rule Score · p. 17

Accessibility

  • Accessibility checks approximate reachability by eroding usable floor space, inflating furniture obstacles, and testing whether targets and valid door openings access a common component.Clearance is 0.20 m for bedrooms, living rooms, and kitchens and 0.10 m for bathrooms; these are conservative geometric proxies, not full human-motion or usability simulation.Source: C Layout Rule Score · p. 17

Room-specific checks

  • Room-specific checks penalize kitchen aisle, table-counter, and appliance work-zone deficits; bathroom fixture placement and mounting issues; bedroom wall embedding; and living-room TV-facing or sofa-overlap issues.The rules include semantic exceptions for legitimate mounting and overlap relationships.Source: C Layout Rule Score · p. 17

D Training variants and outcome-supervised reasoning

Outcome-supervised reasoning GRPO is compared with direct-DSL and other training recipes, with reported gains conditional on successful generation and not isolating reasoning alone. The section also establishes reference-score context and limits of rule-based feedback.

Outcome-supervised reasoning

  • The final reasoning stage trains on free-form reasoning followed by a furniture layout, using only the final layout's scalar reward and no target trace or step-level reasoning reward.The prompt does not prescribe an internal reasoning structure, and a high-scoring answer does not establish that every trace statement is correct.Source: D Training variants and… · p. 18

Training comparison

  • The compared training recipes are not equal-compute estimates of the isolated effect of SFT or reasoning, and the final GRPO stage also uses room-specific completion allowances.The final reasoning model uses 6,144 tokens for kitchens, 4,096 for living rooms, and 3,072 for bathrooms and bedrooms.Source: D Training variants and… · p. 18
  • SFT + GRPO with reasoning has higher retained-layout mean scores than direct-DSL GRPO in kitchens, bathrooms, and bedrooms, while direct-DSL GRPO performs better in living rooms.Reasoning GRPO after SFT exceeds GRPO without SFT in both score summaries for every room type, but kitchen reasoning GRPO excludes 54.0% of samples.Source: D Training variants and… · p. 18

Choice of the final model

  • On 43 held-out kitchens with retained outputs from both branches, reasoning GRPO reaches mean LRS 3.48 versus 1.66 for direct-DSL GRPO; disallowed overlaps fall from 36.6% to 17.1%, while wall-overlap violations rise from 30.2% to 37.0%.Insufficient-clearance layouts fall from 18.9% to 5.8%, and both branches place approximately 11.5 objects per room; these results are conditional on successful generation.Source: D Training variants and… · p. 18

Earlier trace supervision

  • Earlier trace-supervision experiments found reliable trace labels difficult to construct, and one bedroom setup showed that low teacher-forced loss could reflect copying rather than useful reasoning.That result concerned an earlier setup with final coordinates visible in the prompt but masked from SFT loss; it does not establish a limitation of outcome-supervised GRPO.Source: D Training variants and… · p. 19

Limits of outcome feedback

  • LRS cannot certify reasoning correctness, aesthetic coherence, all ergonomic requirements, or the suitability of a particular 3D asset.Ground-truth agreement adds a final-layout signal but does not enumerate all valid alternatives; reward improvements therefore call for diagnostics and visual inspection.Source: D Training variants and… · p. 19

Reference-layout sanity check

  • Annotated AntPlan layouts have overall mean LRS 8.73 and median 10; Architect-Ant’s 8.80 mean on SceneSmith shells concerns a different room population and does not show better-than-professional design quality.Table 12 reports mean ± sample standard deviation; its overall row pools all rooms.Source: D Training variants and… · p. 20, Table 12 · p. 20

E Representation and visual conditioning

The furniture DSL encodes each object as a category and axis-aligned rectangular footprint, supporting object-level edits and deterministic geometric checks. In the untuned base model, adding a structure-only image raises mean LRS across all four room types.

Output format

Representation

  • Each furniture instance is represented by its category and an axis-aligned rectangular footprint, allowing object-level edits and deterministic geometric checks.The representation does not explicitly encode relational triplets, design intent, height, detailed shape, or arbitrary orientation.Source: E Representation and visual conditioning · p. 20

Visual conditioning

  • For the untuned base model, image-conditioned mean LRS is higher in all four room types than without the structure-only image.The comparison holds textual room description and inventory fixed, and does not measure visual conditioning after fine-tuning.Source: E Representation and visual conditioning · p. 20

F From footprints to editable 3D scenes

Architect-Ant’s DSL is rendered into 3D scenes using architectural structure and furniture assets, with some spatial details supplied by deterministic rules. The appendix also compares prompting inputs for the untuned base model.

F From footprints to editable 3D scenes

  • The renderer parses the DSL and architectural structure, builds walls and openings, and instantiates furniture meshes in Blender.Because the DSL does not predict height or yaw, the asset library and deterministic rules supply them; placement rules orient chairs toward tables and televisions toward seating or beds.Source: F From footprints to editable 3D scenes · p. 20
  • Asset fitting selects compatible classes and sizes, while flexible surfaces and modular cabinet runs can be built to target dimensions.Rendering demonstrates a downstream scene-construction path, not independent validation of predicted geometry; evaluation uses original layout geometry rather than presentation meshes or added supports.Source: F From footprints to editable 3D scenes · p. 20
  • Table 13 compares input and prompting choices for the untuned base model using eight samples per held-out room.It reports mean LRS with 95% normal-interval half-widths over retained samples pooled across rooms.Source: Table 13 · p. 21

G Comparison protocols and metric definitions

The comparison evaluates complete furnishing pipelines on shared room shells and reports both layout-based metrics and automated visual preferences. Results favor Architect-Ant on the reported room preferences and mean LRS, with important differences in scope and metric construction.

G Comparison protocols and metric definitions

  • Table 3 compares configured furnishing pipelines on shared shells, including each pipeline’s inventory construction, candidate generation, native refinement, and final selection.Candidate counts and internal search differ, so the comparison measures pipeline outcomes and estimated costs rather than equal candidate or compute budgets.Source: G Comparison protocols and metric… · p. 21
  • The benchmark uses 109 SceneSmith shells, while methods differ in inventory construction, adaptations, and target-shell conditioning.Architect-Ant retrieves room-based furniture requests; SceneSmith uses released scenes, and some 3D-FRONT baselines are not conditioned on target shells, so their out-of-bounds rates also reflect fit to assigned shells.Source: G Comparison protocols and metric… · p. 21
  • SceneSmith uses a planner–designer–critic refinement loop, whereas the other evaluated methods use distinct search, optimization, or sampling procedures.SceneSmith’s loop can run up to 1,000 turns per agent; Holodeck uses a 30-second floor-placement search budget, and LayoutGPT has no subsequent verification-and-revision pass.Source: G Comparison protocols and metric… · pp. 21–22
  • FUN+ measures satisfied room-specific functional slots, while accessibility averages eligible-room fractions and COL+ and OOB capture collision and containment faults.On dining shells, the Gemma-based methods use kitchen slots, unlike SceneSmith and learned baselines; FUN+ therefore does not apply one identical checklist to all methods. Runtime sums for some baselines combine generation and rendering estimates, and the timing scopes do not establish hardware-normalized or equal-compute speedups.Source: G Comparison protocols and metric… · pp. 22–23

H Automated visual preference studies

  • The visual study uses two judges, anonymous A/B renders, and one completed judgment per method per room, with randomized assignments.The prompt asks about naturalness, plausibility, practicality, arrangement, and movement; both completed room studies had no errors or cannot-judge responses.Source: H Automated visual preference studies · p. 23, §H.1 · p. 23

H.1 Shared-shell room comparison

  • Sonnet favored Architect-Ant in 72 room comparisons and SceneSmith in 36, while Kimi favored Architect-Ant in 74 and SceneSmith in 34, with one tie each.The kitchen group includes 27 dining shells mapped to the kitchen setting, alongside 13 kitchen shells.Source: §H.1 · p. 23
  • Across all 109 paired rooms, mean LRS was 8.80 for Architect-Ant and 8.44 for SceneSmith.LRS evaluates layout geometry under rules and also guides training and candidate selection, so this is not an independent quality assessment.Source: §H.1 · p. 23

H.2 Selected whole-house comparison

In a selected whole-house comparison using different floor plans, automated judges preferred Architect-Ant’s rendered collections over the baselines. The authors qualify these preferences as complementary image-based evidence, not a human study or geometric validity test.

H.2 Selected whole-house comparison

  • The 17 Architect-Ant houses were selected from a pool of 100, while each baseline’s 17 were selected from pools of 31.Selection ranked usable functional zones and completeness before tie-breakers, and the independently selected collections have different floor plans.Source: §H.2 · pp. 23–24
  • Each baseline comparison randomly paired the same 17 Architect-Ant houses against its 17-house collection, with randomized A/B order and no reversed-order repeats.The prompt assessed living, circulation, cooking/eating, bathroom, and sleeping functions while allowing open-plan and separated-room layouts.Source: §H.2 · p. 24
  • 88.2% of Sonnet’s selected-house judgments and 94.1% of Kimi’s favored Architect-Ant over the baselines.The totals were 30 of 34 for Sonnet and 32 of 34 for Kimi; the compared houses had different floor plans.Source: Table 14 · p. 24, §H.2 · p. 24

H.3 Judge configuration and interpretation

  • The study presents automated visual preferences, not human-user judgments or a geometric validity test.One presentation order per pair leaves order sensitivity unmeasured, and image visibility can affect whether fixtures are recognized.Source: §H.3 · p. 24

H.2 Selected whole-house comparison

  • Figure 7 illustrates judgments favoring Architect-Ant in two examples and SceneSmith in one, with the judge’s reasons tied to visible layout features.The stated reasons include circulation, room separation, and the visibility or completeness of cooking, dining, and bathroom arrangements.Source: Figure 7 · p. 25

H.4 Room judging prompt

  • The room prompt asks judges to compare visible naturalness, plausibility, practicality, and comfortable movement, without inferring hidden features or precise measurements.The prompt says more furniture is not automatically better and asks judges to focus on arrangement and usability rather than rendering style.Source: §H.4 · p. 26

H.5 Whole-house judging prompt

  • The whole-house prompt evaluates living, circulation, cooking/eating, bathroom, and sleeping functions together with placement, access, and sensible relationships between zones.It allows open-plan or studio layouts and asks judges not to infer unseen appliances, exact measurements, or code compliance.Source: §H.5 · p. 26
Loading 2606.10953v2…