SYNBUILD-3D: Synthesizable and Semantic 3D Data

• SYNBUILD-3D is a synthetic 3D data standard characterized by multi-modal inputs (wireframes, images, point clouds) and detailed semantic annotations.
• It employs a four-step generation pipeline integrating procedural modeling, AI-driven floor plan synthesis, vectorization, and 3D extrusion to produce LoD 4 models.
• The framework extends to applications in building reconstruction, molecule design, scene text rendering, and constraint-based software synthesis.

SYNBUILD-3D denotes a family of standards, datasets, and generative methodologies for synthesizable and semantically rich 3D data across multiple domains, including building modeling, molecule design, scene text rendering, and software synthesis. The SYNBUILD-3D name appears both as a dataset providing multi-modal, high-detail building models and as a canonical label for rigorous synthetic data protocols that integrate geometric and semantic cues to drive state-of-the-art generative modeling, constraint satisfaction, and practical downstream applicability. Its hallmark attributes are the scale of annotated 3D data, explicit multi-modality (e.g., wireframe, images, point clouds), detailed semantic annotation, rigorous geometric encoding, and extensible pipelines for both AI research and applied computational modeling.

1. SYNBUILD-3D for 3D Building Modeling: Dataset Composition and Semantics

SYNBUILD-3D (Mayer et al., 28 Aug 2025) is a large-scale, multi-modal synthetic dataset of over 6.28 million 3D residential buildings at Level of Detail 4 (LoD 4), structured to address the scarcity of annotated, semantically consistent 3D building data. Each instance is parametrically generated from 16,000 unique exterior hulls and 81,223 validated floor plans, with floor plan permutations yielding extensive intra-hull diversity (1–17,160 variants per hull; 87.5% of data from hulls with >1,000 permutations).

Each building comprises three coordinated modalities: (i) a semantically enriched 3D wireframe graph at LoD 4, (ii) corresponding floor plan images with pixel-wise semantic segmentation, and (iii) a LiDAR-like roof point cloud sampled at 50 points/m². LoD 4 per CityGML requires full exterior plus interior geometries, multi-story layouts, and detailed semantic labels for rooms, doors, windows, and walls. In SYNBUILD-3D, wireframes are unified interior/exterior graphs with nodes and edges semantically annotated via a per-node feature matrix, including precise room type IDs (e.g., Living Room=1, Kitchen=3) and structural tags for walls or roof facets. No closed-form mesh parameterization is furnished for roof geometry; connectivity is encoded by adjacency matrices.

Semantic annotation is enforced with consistency constraints: complete room enclosure by walls, mandatory accessibility (each room linked to a door), unique boundary nodes across floors, and global label completeness. The dataset provides ≈395 million semantic labels across 16 element types.

2. Generation Pipeline and Geometric Realization

The SYNBUILD-3D pipeline comprises four main steps: (1) procedural LoD 3 exterior hull generation, (2) AI-driven floor plan synthesis conditioned on the hull footprint, (3) floor plan vectorization into 2D wireframes and semantic graphs, and (4) spatial alignment, extrusion, and stacking to produce final LoD 4 wireframes. The procedural external shell generator is based on Random3dCity with additional hull and extension merging. The RPLAN engine conditions floor plan synthesis on the hull’s footprint geometry and entrance placement; outputs are both RGB and segmentation masks.

Vectorization randomly seeds nodes on structure pixels, builds an initial adjacency graph, then iteratively refines anchor nodes, merges non-anchors, and aligns graph elements to wall directions. The alignment step optimizes translation and scale parameters to best overlay vectorized plans within the hull’s footprint. Floors are extruded into 3D volumes and stacked to conform with the exterior mesh. Roof geometry is handled by sampling points on facets of the procedural mesh, with adjacency relations but without explicit solver-triangulated mesh faces.

3. Modalities, Annotation, and Data Access

SYNBUILD-3D delivers each building as a triplet:

• Modality I: 3D wireframe graph with semantic per-node features. Structural adjacencies, window and door subgraphs, and roof facet graphs are explicitly separated.
• Modality II: RGB floor plans and corresponding segmentation masks at resolutions such as 256×256, with each pixel value assigned a unique semantic ID.
• Modality III: LiDAR-like roof point cloud, sampled uniformly across all roof facets, representing “sensor-style” geometry for learning reconstruction or simulation tasks.

Semantic inheritance is strictly linked—3D nodes inherit room or element labels from originating pixels during vectorization, enforcing cross-modal consistency. The open-source generation pipeline is available, supporting augmentation to new typologies, custom roof geometry, or extended noise models.

4. SYNBUILD-3D as a Formal Data Standard: Extensions to Scene Text and Software Synthesis

As a general data protocol, SYNBUILD-3D also denotes the inclusion of explicit 3D geometric cues (e.g., surface normals, depth, curvature) in synthetic scene text and software synthesis datasets (Hsiung et al., 24 May 2025). The “SYNBUILD-3D standard” for scene text mandates that each datum comprises a natural RGB image (background), composite with rendered 3D-oriented text, a binary mask, and a per-pixel normal map, with rotations precisely encoded via sequential roll, pitch, and yaw. Arc-based distortions and curvature are included to model non-planar layouts.

In software synthesis (Li et al., 24 Jul 2025), SYNBUILD-3D-related approaches formalize complex 3D environments by combining domain-specific languages for object/region declarations and spatial constraints, constraint-solvers (SMT or LLM-refined), and modular stages (from natural language to 3D asset realization). The explicit constraint-tracing and the use of intermediate representations ensure traceability and configurability in highly structured 3D scene generation.

5. Applications, Baselines, and Research Impact

SYNBUILD-3D enables a breadth of tasks: generative modeling of LoD 4 structures conditional on 2D input (e.g., PolyGen, EGNN), semantic and geometric reconstruction from point clouds (e.g., PBWR), multi-modal consistency learning, and simulation-driven architectural research (energy usage, indoor navigation). The dataset’s scale and semantic completeness promote cross-modal generative AI developments. In related 3D molecule and software synthesis fields, the SYNBUILD-3D protocol generalizes to joint discrete-continuous modeling—generating synthesizable structures with explicit retrosynthetic traces or constraint-led 3D environments.

No quantitative model benchmarks are reported for the SYNBUILD-3D building dataset itself, but it supports future evaluation and method development across reconstruction and generation paradigms.

6. Comparative Datasets, Prior Art, and Model Compactness

The concept of synthesizable 3D data has independent precedents in primitive-based building modeling for remote sensing (Li et al., 2019), where recursive spatial partition, synthetic height fields, and primitive instance segmentation are fused via deep models and rigid point cloud fitting. That line of work achieves IoU>80% for 3D reconstruction and orders-of-magnitude reduction in model size relative to mesh baselines. In synthesizable molecule generation (Rekesh et al., 16 Jul 2025, Li et al., 2022), joint modeling of building-block graphs, reaction templates, and 3D atomic coordinates is framed as a diffusion/flow process with explicit mask-based action space restriction for synthesizability, supporting unconditional, conditional, and “inpainting” (fragment-linking) tasks with high validity and synthetic accessibility.

The SYNBUILD-3D philosophy—fusing high granularity in geometry, semantics, and traceability—represents a convergence of best practices across computer vision, computational chemistry, and AI-driven software synthesis.

7. Future Directions and Limitations

For 3D buildings, pipeline extensibility includes augmenting the dataset with novel structural types, more complex roof geometries, and diverse sensor noise models. For scene text and multimodal 3D-aware generative tasks, proposals for future work include deeper curvature encoding, per-pixel depth, and additional mapping (illumination, roughness) to enrich geometric context. Practical deployment in software and molecule synthesis domains remains influenced by extant template rule sets and domain constraints; improvements depend on integrating reaction yield predictors, automating anchor specification, and refining pocket-conditioning protocols.

A plausible implication of SYNBUILD-3D’s structure is that it fosters a new generation of data-driven, semantically grounded models for geometric synthesis, accommodating strict semantic-geometric alignment, traceable constraint satisfaction, and multi-modal research workflows (Mayer et al., 28 Aug 2025, Hsiung et al., 24 May 2025, Li et al., 24 Jul 2025, Li et al., 2022, Rekesh et al., 16 Jul 2025, Li et al., 2019).