BIM-Native Tokenization for Constraint-Aware Room Layout Synthesis

SBM generates functionally correct and semantically coherent layouts given a room envelope. Each row shows a different room type. — **Small Building Model (SBM)** is an encoder-decoder Transformer that generates functionally correct and semantically coherent layouts given a room envelope. Each row shows a different room type. Our approach outperforms general-purpose LLMs/VLMs and domain-specific methods.

Demo Video

Abstract

We present a BIM-native tokenization for room-level layout synthesis in Building Information Modeling (BIM) scenes. The core contribution is representational: we encode each room as a sequence of BIM-Token Bundles, realized as columns of a sparse attribute–feature matrix that unifies categorical and continuous attributes of walls, openings, and entities under wall-referenced (translation/scale-invariant) coordinates.

A mixed-type embedding module produces a unified token vector from this matrix; a single Transformer backbone is then trained in two modes: encoder-only for room embeddings and retrieval, and encoder–decoder for autoregressive entity placement, which we call Data-Driven Entity Prediction (DDEP). On a controlled same-data benchmark with shared ontology and evaluation harness, DDEP outperforms ATISS and BLT baselines bridged into our representation, with ablations identifying joint continuous-feature embedding and entity ordering as primary drivers. Encoder embeddings cluster rooms by type more tightly than large general-purpose text encoders, which in turn retain an edge on within-type ranking. We frame this work as evidence that modestly sized, domain-specific sequence models over well-designed BIM tokenizations are a useful primitive for constraint-aware spatial generation, complementary to general-purpose LLMs/VLMs which we also benchmark.

How Does It Work?

Our goal is to learn a unified representation for BIM scenes that enables both retrieval (finding similar rooms) and generation (placing furniture, doors, and fixtures). SBM takes as input a room envelope—walls, doors, windows—and produces either room embeddings or a complete layout with placed entities.

SBM Architecture Overview — **Model Overview.** (a) BIM data extraction and assembly into a discrete set of token bundles. (b) SBM encoder stack processes the tokenized feature-attribute matrix and outputs a room representation. (c) SBM decoder stack consumes the room representation as memory to the cross-attention layers and the room entities as inputs, trained on next token prediction. (d) Downstream use cases. Retrieval and unguided layout generation (DDEP) are evaluated in the paper; the user-guided path through an agentic layer is a prospective integration that BIM-native output tokens enable but that we do not evaluate.

BIM-Token Bundles

We introduce a normalized hierarchical tokenizer that encodes room topology, entity attributes, wall-referenced geometry, and relational structure into compact token bundles realized as rows of a sparse attribute-feature matrix.

Mixed-Type Embeddings

A unified embedding module handles both categorical and continuous features, learning joint representations that preserve the compositional structure of architectural elements.

Dual-Mode Backbone

A single Transformer supports encoder-only room embeddings for retrieval and encoder-decoder DDEP for autoregressive layout generation.

Mixed-type embedding, step by step. Each feature row of the sparse attribute-feature matrix has its own embedder: a lookup table for categorical features, a small MLP for continuous ones, and a joint embedding for the grouped edge endpoints. Masked cells become exact zero vectors. A gated sum then adds the features of each token down its column, never across tokens, so S tokens go in and S vectors come out, one per BIM element, ready for the Transformer encoder.

Encoder-Only Pathway

The encoder-only pathway produces high-fidelity room embeddings that capture geometric, topological, and semantic properties. These embeddings cluster reliably by room type and topology, achieving NMI: 0.726 (2.0× better than text embedding baselines) and near-perfect type-constrained triplet accuracy (99.9%).

Data-Driven Entity Prediction (DDEP)

In DDEP mode, the encoder-decoder pipeline autoregressively places room entities (props and casework) conditioned on the room envelope. DDEP produces functionally sound layouts with:

Near-complete inventory satisfaction — placing required furniture and fixtures
State-of-the-art navigability — ensuring walkable paths through the room
Lowest violation rates — fewer collisions and boundary violations than LLM/VLM baselines

Evaluation

We evaluate SBM under two complementary protocols, exactly as reported in the EUVIP 2026 paper. The controlled same-data benchmark is the lead result: every method is trained from scratch on one frozen split with a shared ontology, canonical furniture dimensions and a single evaluation harness, which isolates architectural effects from data scale. The deployment-style benchmark compares production DDEP against zero-shot frontier LLMs/VLMs and domain-specific systems. Metrics are Cov. (inventory coverage ↑), Nav (door-to-target navigability ↑), SR (reachability ↑), DF (detour factor ↓) and OC (overlap–clearance ↓); they must be read jointly, since a sparse room can score well on OC while failing coverage and reachability.

Controlled same-data benchmark

Method	Cov. ↑	Nav ↑	SR ↑	DF ↓	OC ↓
ATISS (same-data)	40.8 ± 23.6	−30.3 ± 24.0	0.03	0.97	0.5
BLT (same-data)	17.8 ± 20.4	−18.6 ± 42.7	0.12	0.88	35.7
DDEP (ours)	54.9 ± 38.3	14.6 ± 56.0	0.36	0.60	13.4
DDEP ablations
Wall-only	50.8	9.6	0.32	0.63	12.7
No continuous groups	42.6	0.7	0.25	0.69	13.4
No edge coordinates	56.7	14.4	0.35	0.60	14.5
Order: t first	42.6	4.1	0.29	0.70	10.6
Order: edge, t desc.	46.5	7.8	0.31	0.66	13.4

Shared 1,225-room geometry-valid test subset. ATISS's near-zero OC reflects sparse outputs that avoid collisions while leaving almost all targets unreachable; BLT shows the opposite failure mode. Removing the joint continuous-group embedding is the single most damaging change, which is the paper's central evidence that the representation — not model scale — drives performance. ATISS and BLT outputs are bridged into our wall-referenced format; that bridge is applied to the baselines but not to DDEP, so part of the gap may reflect conversion loss.

Deployment-style benchmark vs. frontier models

Method	Cov. ↑	Nav ↑	OC ↓	Latency (s) ↓
Text-only LLMs
Claude Opus 4.6	58.9	−0.8	11.3	15.6
Claude Sonnet 4.6	64.9	0.3	10.5	81.2
Claude Haiku 4.5	55.3	4.4	15.2	7.0
GPT-5.2 ^†	47.6	16.8	13.0	55.1
Gemini 3.1 Pro ^‡	63.7	20.5	9.2	67.1
Codex 5.3	68.3	18.9	9.8	20.8
Qwen 3.5	64.6	3.1	13.6	19.0
Vision-language models (floorplan-conditioned)
Claude Opus 4.6	61.9	55.5	8.2	15.3
Claude Sonnet 4.6	65.6	34.6	11.0	14.8
Claude Haiku 4.5	60.7	25.2	16.2	6.9
GPT-5.2 ^†	57.0	34.9	12.1	101.7
Gemini 3.1 Pro ^§	86.5	47.8	2.2	–
Qwen 3.5 ^†	60.7	27.1	12.8	84.1
Domain-specific methods
LayoutVLM	–	40.3	3.8	145.0
FlairGPT	46.6	28.2	3.9	1133.6
DDEP (ours)	98.1	79.2	1.9	3.18

50-room held-out set. All models receive the same tuned prompt with an explicit coordinate system, wall-condition guidance, a room-type-filtered entity catalog for information parity with DDEP, and a matched example of the expected output format — so the gap is not an artifact of naive prompting. It remains a deployment floor rather than an architectural ceiling: each model is run once per room with no best-of-k or chain-of-thought, and no open model is fine-tuned in-domain. ^†evaluated on a subset (parse success <70%). ^‡37/50 rooms. ^§7/50 rooms, unreliable.

Specification

Near-Complete Inventory Satisfaction

Navigability

State-of-the-Art Walkable Paths

Overlap

Lowest Collision Rate

Clearance

Lowest Boundary Violations

Our results demonstrate that modestly sized, domain-specific sequence models over well-designed BIM tokenizations can outperform much larger general-purpose systems on layout quality. DDEP delivers functionally sound layouts that satisfy design specifications while maintaining geometric validity and constraint satisfaction.

Room Embeddings

In encoder-only mode, SBM produces high-fidelity room embeddings that capture geometric, topological, and semantic properties. We compare against strong text embedding baselines (E5-Large-v2, BGE, GTE) on clustering quality and type-constrained retrieval.

UMAP visualization comparing SBM and E5 embeddings — **UMAP visualization of room embeddings** colored by room type category. SBM embeddings (left, NMI 0.726) form well-separated clusters with distinct boundaries between room types, while E5-Large-v2 embeddings (right, NMI 0.371) show intermingled clusters with blurred boundaries. The roughly 2× higher NMI reflects SBM's specialization in geometric and spatial structure rather than lexical similarity.

NMI Score

0.726

2.0× better than text baselines

Triplet Accuracy

99.9%

Type-constrained retrieval

Entity Overlap

77.0%

Retrieved rooms share similar furniture

Layout Generation Examples

SBM autoregressively places entities across diverse room types

Gallery

DDEP qualitative results across different room types

Key Contributions

1

Normalized Hierarchical Tokenization

A BIM tokenization scheme with a mixed-type embedding module that unifies categorical and continuous architectural features.

2

Unified Transformer Backbone

SBM supports both retrieval (encoder-only) and generative layout synthesis (encoder-decoder DDEP) in a single architecture.

3

Comprehensive Evaluation

Rigorous benchmarks on a professional BIM dataset demonstrating superior scene coherence, geometric validity, and constraint satisfaction.

4

Domain-Specific Excellence

Modestly sized, domain-specific models over well-designed tokenizations outperform much larger general-purpose systems on layout quality.

BibTeX

@inproceedings{sbm2026euvip,
  title={BIM-Native Tokenization for Constraint-Aware Room Layout Synthesis},
  author={Ladron de Guevara, Manuel and Rhee, Jinmo and Bidgoli, Ardavan and Razgaitis, Vaidas and Bergin, Michael},
  booktitle={Proceedings of the 14th European Conference on Visual Information Processing (EUVIP)},
  year={2026},
  eprint={2512.04832},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}