RenoLayout: Unified Floorplan Renovation and Furniture Layout for Constraint-Aware Residential Design

Jingguo Liu; Yuanqing Liu; Chong Zhang; Yue Yang

cs.CLOct 15, 2026

RenoLayout: Unified Floorplan Renovation and Furniture Layout for Constraint-Aware Residential Design

Authors: Jingguo Liu, Yuanqing Liu, Chong Zhang, Yue Yang

Organizations: Beike

Abstract

Residential design aims to transform existing homes into functional living spaces while satisfying diverse design preferences and physical constraints. However, existing floorplan generation methods typically assume empty building boundaries as input, whereas renovation requires reasoning over existing floorplans. Furthermore, floorplan renovation and furniture layout are commonly addressed as separate tasks, overlooking their inherent spatial dependency. Consequently, practical residential design must preserve immutable structural components while transforming editable spatial elements. It must also optimize furniture placement under design objectives and physical constraints. To address these challenges, we formulate residential design as a design-point-guided conditional generation task and propose RenoLayout, a framework that translates a pre-design residential state together with a design-point set into an executable Geometry-as-Code program. To enable controllable supervision, we construct a Design Map comprising 123 fine-grained design points covering structural renovation and furniture layout. We further introduce Difference-Grounded Design-Point Annotation to derive design-point sets automatically from before--after design pairs with VLM-based consistency verification. RenoLayout is trained using a three-stage CPT--SFT--RL pipeline, progressively improving spatial reasoning, instruction following, and constraint compliance. We also introduce the RenoLayout Benchmark, a set of 200 real-world before--after renovation cases with complete furniture layouts. To the best of our knowledge, it is the first benchmark for unified residential design. These cases include design-critical structural and building-service constraints.

Figures & tables

Figure 1: Comparison of existing task-specific paradigms with RenoLayout. Existing methods treat floorplan generation separately from furniture layout, whereas RenoLayout unifies constraint-aware renovation with full-home furnishing.

Task setting Representative works	Pre-design state	Post-design solution	Full-home furnishing	Load-bearing walls	Doors & windows	Pipes & drainage
Floorplan generation House-GAN ( Nauata et al. 2020 ) ; Yin et al. ( Yin et al. 2025 ) ; HouseMind ( Qin, Weber, and Lu 2026 )	✗	✓	✗	✗	✗	✗
Scene layout generation LayoutGPT ( Feng et al. 2023 ) ; OptiScene ( Yang et al. 2025b ) ; HoGLayout ( Jiang et al. 2026 ) ; LaviGen ( Feng et al. 2026 )	✗	✗	✓	✗	✓	✗
Joint floorplan-layout generation HouseDiffusion ( Shabani, Hosseini, and Furukawa 2023 ) ; Nestwork ( Miao, Cao, and Zhuang 2026 )	✗	✓	✓	✗	✓	✗
Unified residential design RenoLayout (Ours)	✓	✓	✓	✓	✓	✓

Table 1: Comparison of data annotations in residential design. Unlike existing settings, RenoLayout provides paired pre-design states with post-design solutions, complemented by full-home furnishing as well as design-critical annotations for structures and building services.

Figure 2: Frequency of furniture-layout (top) and structural-renovation (bottom) design points in the training (blue) and test (red) splits. Frequency is the occurrence count normalized by the total number of samples per split; design points are ordered by descending training-set frequency.

Figure 3: Overview of RenoLayout. Left: paired designs are converted into difference-grounded design-point sets and verified through multi-VLM consistency auditing. Middle: the pre-design state is serialized as Geometry-as-Code and combined with design-point prompts, enabling Qwen3 to first produce a graph-based spatial plan and then generate executable structural-renovation and furniture-layout actions through the CPT–SFT–RL training pipeline.

Category	Total	Avg./case	Std.	Min	Max
Renovation	875	4.38	2.33	1	13
Layout	2,177	10.88	3.40	0	23
Overall	3,052	15.26	4.71	2	31

Table 2: Per-case statistics and total counts of structural-renovation and furniture-layout design-point annotations across the 200-case RenoLayout Benchmark.

Settings			Overall Metrics		VLM-HS by Evaluator
CPT	Box	CoT	Recall@1 $\uparrow$	Pass@1 $\uparrow$	Seed	GLM	GPT	Gemini
✗	✗	✗	80.92%	27.00%	5.12	4.97	5.60	5.07
✗	✓	✗	81.13%	38.50%	5.86	5.85	6.39	5.77
✗	✓	✓	84.47%	42.00%	5.98	5.86	6.46	5.78
✓	✗	✓	88.65%	31.50%	6.32	6.01	6.69	6.19
✓	✓	✗	86.37%	45.00%	6.61	6.60	7.10	6.40

Table 3: Feature ablation of CPT, bounding-box encoding (Box), and graph-planning CoT. All metrics except VLM-HS (0–10) are percentages.

Figure 4: Qualitative effects of bounding boxes on local placement and CoT on spatial planning.

Method	Overall Metrics		VLM-HS by Evaluator
	Recall@1 $\uparrow$	Pass@1 $\uparrow$	Seed	GLM	GPT	Gemini
RenoLayout-SFT	88.67%	49.00%	6.82	6.75	7.22	6.45
+ DPO	89.37%	55.50%	6.92	6.90	7.36	6.64
+ GDPO	90.89%	53.00%	6.87	6.81	7.20	6.77
+ DPO + GDPO	91.63 %	61.00%	7.26	7.32	7.72	7.22
+ GDPO + DPO	90.42%	61.50 %	7.17	7.23	7.70	7.16

Table 4: RL post-training ablation of DPO and GDPO. All metrics except VLM-HS (0–10) are percentages.

Method	Designer 1	Designer 2	Designer 3	Mean $\uparrow$	Std. $\downarrow$
RenoLayout	5.50	5.59	5.19	5.43	0.17
Ground Truth (GT)	7.06	7.35	6.58	7.00	0.32

Table 5: Professional Approval Score (PAS; 0–10).

Figure 5: Visual comparison with LayoutGPT. Red arrows indicate furniture accessibility issues, and blue arrows indicate furniture misalignment with walls. RenoLayout produces more accessible and better-aligned room layouts in these examples.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Structural-Renovation Design Points (52)
No.	Design Point	No.	Design Point	No.	Design Point
1	Bedroom Storage	19	Remove Entryway	36	4BR to 3BR
2	Door/Window Adjustment	20	2BR to 3BR	37	Master Suite
3	Wet-Dry Separation (2-Zone)	21	Add Study in Living Room	38	Wet-Dry Separation (3-Zone)
4	Convert to Utility Area	22	Enlarge Bedroom	39	Kitchen-Dining Integration
5	Entryway Storage	23	3BR to 2BR	40	Corridor Storage

Appendix

Table 7: Complete Design Map containing 52 structural-renovation and 71 furniture-layout design points. Entries are arranged in three parallel numbered blocks for compact presentation.

Dimension	Metric	Range	Description and Key Indicators
Design-point fulfillment	Recall@1	0–100%	Mean case-level proportion of requested design points fulfilled in the first generated outputs, averaged across all 200 cases by a rule-based evaluator.
Constraint compliance	Pass@1	0–100%	Percentage of one-sample generated results that pass all 22 checks across geometric validity, structural compliance, functional integrity, and collision-free placement.
Constraint compliance	Geometric Validity ( $\mathcal{P}_{\mathrm{geom}}$ )	0–100%	Percentage of generated designs with well-formed polygons, valid references, and complete required fields.
Constraint compliance	Structural Compliance ( $\mathcal{P}_{\mathrm{struct}}$ )	0–100%	Percentage of generated designs that preserve load-bearing walls, exterior envelopes, and shafts while maintaining room connectivity through valid doors.
Constraint compliance	Functional Integrity ( $\mathcal{P}_{\mathrm{func}}$ )	0–100%	Percentage of generated designs that retain fixed infrastructure and essential fixtures without damage or omission.
Constraint compliance	Collision-Free Placement ( $\mathcal{P}_{\mathrm{coll}}$ )	0–100%	Percentage of generated designs without furniture overlap or forbidden-zone violations.

Appendix

Table 8: Overview of the unified evaluation framework.

Submetric	Checks	Violation Flags and Annotations
Geometric Validity ( $\mathcal{P}_{\mathrm{geom}}$ )	5	polygon_invalid : malformed polygon geometry; polygon_self_intersect : polygon edges intersect one another; invalid_reference : an entity references a nonexistent or invalid object; missing_required_field : a mandatory geometry or semantic field is absent; wall_endpoint_in_opening : a wall endpoint lies inside a door or window opening
Structural Compliance ( $\mathcal{P}_{\mathrm{struct}}$ )	11	load_bearing_wall_removed : an existing load-bearing wall is deleted; load_bearing_wall_moved : a load-bearing wall is displaced; load_bearing_wall_thinned : load-bearing wall thickness is reduced; load_bearing_wall_new_opening : a new opening is cut into a load-bearing wall; load_bearing_wall_opening_enlarged : an existing load-bearing-wall opening is enlarged; exterior_wall_altered : the exterior envelope geometry is modified; facade_new_window : a new window is introduced on the facade; wall_crosses_shaft : a wall intersects a protected service shaft; opening_exceeds_wall : an opening extends beyond its supporting wall segment; no_entry_door : the dwelling has no valid entrance door; room_no_door : a room has no valid doorway
Functional Integrity ( $\mathcal{P}_{\mathrm{func}}$ )	4	kitchen_flue_altered : the fixed kitchen exhaust flue is modified; gas_pipe_altered : the fixed gas-pipe location or geometry is modified; entry_door_altered : the fixed entrance door is modified; bathroom_missing_fixture : an essential bathroom fixture is absent
Collision-Free Placement ( $\mathcal{P}_{\mathrm{coll}}$ )	2	item_overlap : furniture or other placed items overlap; furniture_forbidden_position : furniture occupies a prohibited wall, opening, or clearance zone
Total	22	Pass@1 requires that none of the 22 violation flags be triggered.

Appendix

Table 9: Exact rule composition of the four constraint-compliance submetrics, with a brief annotation for each violation flag.

Figure 6: Representative failure cases involving irregular boundaries and severely constrained spaces.

Prompt component	Instruction content
Role and task	You are a senior residential design reviewer. Evaluate the post-design plan dimension by dimension. Assess subjective design quality that geometry cannot fully determine in Dimensions 1–5. Assess physical compliance together with hard defects only in Dimension 6.
Design context and design points	When both images are available, they depict the same dwelling in two states. The first image shows the original plan, whereas the second presents the post-design plan under review. Score only the post-design plan; the original plan is provided solely to explain what changed and why. All scores reflect the absolute quality of the post-design solution, not improvement relative to the original plan. The listed design points are the required design objectives. Dimensions 1–4 evaluate absolute zoning, layout, circulation, and amenity quality. They must not mechanically reward or penalize compliance with a design point. However, consequences necessarily caused by correctly implementing a required design point must not be treated as defects. Dimension 5 alone evaluates whether design points were implemented. It also assesses whether their implementation is natural and coherent.
Objective physical-inspection facts	[Objective physical-inspection facts detected by the deterministic engine; the primary basis for Dimension 6.] {Insert the engine’s P0/P1 fact list here; if there are no hits, write: No P0/P1-level physical issues detected.} Cross-check the list against the post-design image. Clearly inconsistent entries may be treated as false detections and downweighted. Do not fabricate physical issues beyond the list. The fact list is used only for Dimension 6 and must not adjust Dimensions 1–5.

Appendix

Table 12: VLM-HS scoring prompt, Part I: role, design context, and physical-inspection input.

Prompt component	Instruction content
Dimension 1: Functional Zoning	Evaluate whether active/quiet, public/private, and wet/dry zones are clearly layered with limited mutual interference. Check whether bedrooms form an independent cluster separated from living/dining/kitchen spaces. Assess whether the plan progresses from entry to public/social space and then reaches private sleeping space. Determine whether bathrooms serve either the sleeping area or the public area rather than remaining isolated centrally. Finally, assess whether living, dining, and kitchen areas form a coherent social zone. This dimension evaluates zoning logic, not missing furniture or reachability.
Dimension 2: Spatial Layout	Evaluate whether proportions are comfortable and balanced. Determine whether the circulation area is economical and the primary/secondary hierarchy is appropriate. Check whether major rooms are roughly within a 1:1–1:2 aspect-ratio range. Assess whether corridors are compact rather than long, dark, or wasteful. Determine whether the master bedroom is larger than secondary bedrooms. The living room should be the largest or second-largest public space. Finally, check whether obvious fragmentary space exists. This dimension evaluates proportion and layout rather than exact floor area.
Dimension 3: Circulation Quality	Assuming rooms are reachable with no must-pass-through-bedroom defect, evaluate the smoothness and quality of daily circulation. Check whether the entry provides a foyer or dust-buffer zone. Assess whether kitchen and dining areas are adjacent with a short housework path. Determine whether the nighttime bedroom-to-bathroom path is short, quiet, and minimally exposed to public activity. This dimension evaluates circulation quality, not objective reachability.

Appendix

Table 13: VLM-HS scoring prompt, Part II: definitions of Dimensions 1–3.

Prompt component	Instruction content
Dimension 4: Amenity Configuration	Evaluate whether the plan provides hallmark configurations of quality living. Check for a master suite with a dedicated bathroom or dressing/storage area. Also assess whether the plan includes a distinct entry foyer with systematic storage. Where multiple bathrooms exist, evaluate owner/guest separation. Finally, assess whether an open, connected living/dining area supports social use. Evaluate only items visible in the image; skip unclear evidence rather than fabricating it.
Dimension 5: Design-Point Implementation	Evaluate the listed design points in terms of both fulfillment and implementation quality. Fulfillment asks whether each required design point is visibly realized without omission, distortion, or reversal. Implementation quality considers location, scale, and spatial relationships. These aspects should appear natural and coherent with the overall plan rather than forced or achieved at the expense of visible living quality. This dimension must be supported by a mandatory item-by-item design-point audit.
Dimension 6: Physical Compliance	Evaluate objective physical compliance and defects. Use the deterministic P0/P1 fact list as the primary basis, cross-checked against the post-design image. P0 defects are the primary ordering signal and must be strongly penalized; P1 defects provide only finer differentiation within a severity band. Do not invent issues beyond the fact list.
Scope boundary	The following objective items are evaluated only in Dimension 6: room reachability; furniture overlap, through-wall placement, or door blockage; missing essential furniture; natural daylight; must-pass-through-bedroom circulation; direct kitchen-to-toilet-bathroom connection; door or passage width; and load-bearing/exterior-wall modifications. Dimensions 1–5 evaluate only subjective design quality within their stated scope. Design-point implementation is evaluated primarily in Dimension 5, while correctly required design-point consequences must not be mechanically treated as defects in Dimensions 1–4.

Appendix

Table 14: VLM-HS scoring prompt, Part III: definitions of Dimensions 4–6 and scope boundaries.

Prompt component	Instruction content
Deduction/bonus checklist	Functional Zoning Deductions: public functions interrupted by bedrooms or bathrooms; rooms arranged as a single-axis sequence without clustering; entry directly entering or facing the bedroom zone; a bedroom sandwiched between public spaces; or a centrally isolated bathroom. Bonuses: bedrooms grouped independently behind a corridor or foyer, and a clear entry $\rightarrow$ public $\rightarrow$ private progression. Circulation Quality Deductions: no foyer buffer; kitchen and dining separated by a detour; nighttime bedroom-to-bathroom circulation crossing public activity space; crossing or strongly interfering daily paths; or an obviously circuitous route from entry to major functions. Bonuses: an entry buffer, adjacent kitchen/dining spaces, and a short private nighttime path. Amenity Configuration Deductions: an ordinary master bedroom with neither dedicated bathroom nor dressing/storage area; no entry zone or shoe-storage location; scattered storage; multiple bathrooms without owner/guest separation; or a living/dining area fragmented by circulation. Bonuses: a master suite, foyer plus systematic storage, owner/guest-separated bathrooms, and an open connected living/dining area. Design-Point Implementation Audit every listed design point individually. A point is Implemented only when its required action produces the intended outcome. Partially implemented indicates that the direction is present but a key action, scale, or placement is missing. Not implemented indicates no visible evidence. Reversely implemented indicates an outcome contrary to the requirement. Score this dimension according to fulfillment as well as implementation quality. Consider the resulting placement together with its spatial relationship. Both should appear natural and coherent.
Scoring-band anchors	Spatial Layout —9–10: balanced major-room proportions, no elongated forms, no or only a very short dedicated corridor, appropriate room hierarchy, with no fragmentary space; 7–8: most rooms proportioned appropriately, moderate corridor share, with at most minor local narrowness; 6–7: one or two clear scale problems; 5–6: multiple scale problems, large fragmentary space, excessive corridor share, or reversed master/secondary hierarchy; 0–4: severely unbalanced layout. Physical Compliance —9–10: $P0=0$ , $P1\leq 2$ ; 8–9: $P0=0$ , $P1=3$ – $6$ ; 7–8: $P0=0$ , $P1\geq 7$ ; 5.5–7: $P0=1$ ; 3–5.5: $P0=2$ ; 0–3: $P0\geq 3$ . P1 issues may gradually lower the score within a band, but must not outweigh P0 defects.
Hidden design-defect checklist	Living/dining space traversed by circulation and turned into a passage; a severely imbalanced frontage-to-depth ratio creating a strip-like plan; excessive room subdivision causing cramped rooms or fragmentary blocks; a main public space smaller than a secondary room; a bathroom door facing the entry, dining table, or bed; or an open kitchen venting directly toward living/dining space without any partition.

Appendix

Table 15: VLM-HS scoring prompt, Part IV: checklists and scoring anchors.

Prompt component	Instruction content
Scoring procedure	A. Dimensions 1–5: First inspect the relevant deduction/bonus checklist together with the hidden-defect checklist. Record only visible hits with their locations. Then assign a score based on the number and severity of hits rather than a vague overall impression: four or more deductions, or one very severe deduction, normally yields 5–6; three deductions yields about 6; two deductions yields 7–8; at most one minor deduction yields about 8; no deduction yields 8–9. Multiple visible bonuses with no deduction may yield 10. Scores of 6 or below must name the specific deduction and location; scores of 8 or above must be supported by no obvious deduction. For Dimension 5, audit every design-point ID separately. Each item must be labeled Implemented , Partially implemented , Not implemented , or Reversely implemented , with image-grounded evidence. Do not merge multiple IDs using statements such as “most” or “all.” A design point is implemented only when its key action and intended result are both visible; unclear evidence is labeled partially implemented. The dimension score reflects both fulfillment and implementation quality, but is not mechanically fixed by the proportion of item statuses. B. Dimension 6: Use the P0/P1 fact list as the primary basis. Cross-check obvious inconsistencies against the image, do not fabricate additional issues, and assign the score using the P0/P1 anchors. A true $P0$ must substantially lower the score: $P0=0$ normally implies at least 7, $P0=1$ cannot exceed 7, and $P0\geq 2$ cannot exceed 5.5.
Output format	Output only one strict JSON object, with no draft reasoning or other text. The design_point_checks array is mandatory: every supplied design-point ID must occur exactly once; output an empty array if no design points are supplied. Reasons must be concise, image-grounded, and consistent with the item-level audit. Each reason states at most two deductions or bonuses with locations; vague boilerplate is prohibited. If a dimension cannot be judged from the rendering, use score $-1$ and explain why; score 0 is reserved for an extremely poor but judgeable plan. Do not output an overall score. The JSON has the form: {"design_point_checks": [{"id": "L2-1", "status": "Implemented", "evidence": "..."}], "dimensions": [{"dimension": "Functional Zoning", "score": <0--10>, "reason": "..."}, {"dimension": "Spatial Layout", ...}, {"dimension": "Circulation Quality", ...}, {"dimension": "Amenity Configuration", ...}, {"dimension": "Design-Point Implementation", ...}, {"dimension": "Physical Compliance", ...}]}

Appendix

Table 16: VLM-HS scoring prompt, Part V: scoring procedure and output format.

Constrained Generative Modeling

Constrained generative modeling develops methods that produce samples satisfying structural, physical, semantic, or user-defined requirements.

See all 185 papers

Jul 26, 2026

PlanCraft: Sketch, Refine, and Furnish for Architect-Inspired Progressive 3D Residential Scene Generation

Two structural insights have been overlooked in automated residential floor plan generation. First, design is inherently progressive. Architects begin with rough strokes and refine them over time, whereas existing methods typically require their conditioning representation to be fully specified before generation, a fundamental mismatch with how design actually works. Second, the 2D floor plan is not an optional intermediate but an irreplaceable spatial contract. Once room boundaries, doors, and windows are fixed, furnishing reduces from open-ended spatial reasoning to bounded constraint satisfaction. Bypassing this contract, as existing 3D systems do by delegating layout to language models, yields overlapping rooms and implausible proportions; directly calling general-purpose language models likewise produces geometrically invalid layouts. Guided by these insights, we present PlanCraft. SketchPlan supplies the missing training signal by replaying the architect's drawing process on 80K real floor plans, producing partial sketches at every completeness level. PlanCraft-Diff progressively sharpens an incomplete sketch into a geometrically precise, vectorizable floor plan through a coarse-to-fine strategy. With the spatial contract established, PlanCraft-Agent then furnishes the scene within well-defined room boundaries. Experiments show that PlanCraft achieves a 61.1% lower FID than the best existing 2D method and surpasses existing 3D systems by 15 points in expert-rated spatial rationality, with a sketch at only 25% completion already outperforming all fully specified baselines.

Pengyu Zeng, Yuqin Dai, Jun Yin +7

Tsinghua University

3D Scene Generation 3D Scene Layout Generation Floor Plan Generation Indoor Scene Generation

Ask PDF

Jun 9, 2026

Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

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.

Fedor Rodionov, Aleksandar Cvejic, Michael Birsak +2

King Abdullah University of Science and Technology (KAUST), Saudi Arabia · Miami University, United States of America

Procedural Content Generation Indoor Scene Generation

Ask PDF

Jul 9, 2026

GRE-Diff: Gaussian Room Embeddings for Structured Layout Diffusion

Designing functional and aesthetically coherent floor plans requires exploring a vast space of possible room arrangements, a task that quickly becomes overwhelming for human designers. In this paper, we propose GRE-Diff, a controllable and interactive diffusion-based framework that automates the creation and editing of apartment floor plans under user-specified constraints. By combining AI-generated suggestions with real-time, human-in-the-loop editing, the system enables users to specify room types, room counts, boundary shapes, and editing operations through LLM-parsed instructions or GUI-based interaction. It then generates a diverse set of plausible and well-structured designs for refinement. At the core of our approach is Gaussian Room Embedding (GRE), a continuous latent representation that models each room as a spatial Gaussian distribution capturing its location and extent. Extensive experiments on the RPLAN dataset show that GRE-Diff produces high-quality, constraint-aware, and editable polygonal layouts, offering a practical step toward bridging AI-driven automation and human creativity in spatial design.

Jing Wang, Haoran Xiong, Zihao Yan +2

Guangdong Provincial Key Laboratory of Visual Media and Multidimensional Intelligence, CSSE, Shenzhen University, China · School of Computer Science, University of Guelph, Canada

Floor Plan Generation Indoor Scene Generation Human-AI Co-Creation Diffusion Models

Ask PDF

Abstract

Constrained Generative Modeling

Explore similar work

PlanCraft: Sketch, Refine, and Furnish for Architect-Inspired Progressive 3D Residential Scene Generation

Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

GRE-Diff: Gaussian Room Embeddings for Structured Layout Diffusion