Heritage-Aware Generative AI Workflow for Islamic Geometry in Interiors

Abstract

Recent text to image systems can synthesize Islamic heritage elements with high visual fidelity, but their outputs rarely translate into fabricable geometry or integrate into interiors without substantial redrawing. We present an end-to-end workflow that links historically grounded precedent retrieval, controllable tileable generation, semantic segmentation and vectorization, and geometry-aware mapping into Computer-Aided Design (CAD) environments. Contributions include the following: (i) a license-audited dataset schema and a retrieval classifier for common Islamic motif families and architectural elements; (ii) precedent retrieval via a ResNet 50 and Vision Transformer (ViT) embedding pipeline; (iii) a Low-Rank Adaptation (LoRA) tuned diffusion model that generates tileable motifs with motif/region controls; (iv) a raster-to-vector pipeline that enforces curve closure and minimum feature widths for CNC/laser fabrication; and (v) a rubric and domain metrics (symmetry coherence, seam/tileability error, spline closure and junction valence, UV distortion, feature width compliance) that quantify “depth of integration” beyond surface texture. Quantitative metrics and blinded expert ratings compare the workflow against strong parametric baselines, while scripts translate images to fabrication-ready vectors/solids across walls, ceilings, partitions, floors, and furniture. Cultural safeguards cover calligraphy handling, regional balance audits, and provenance/credit. The workflow advances heritage-aware generative design by carrying imagery across the last mile into buildable detail and by providing practical checklists for adoption in interior architecture and conservation.

1. Introduction

Interior designers are increasingly asked to deliver spaces that celebrate Islamic heritage while meeting contemporary performance and fabrication constraints [1]. Rapid progress in generative models has made high-fidelity ornament synthesis accessible to non-experts [2,3], yet most outputs remain image-bound and require substantial manual redrawing before integration as buildable detail [4,5]. The result is a widening gap between compelling visuals and production-ready interior elements, with recurring risks of superficial or tokenistic application that neglect structural seams, service zones, and cultural signifiers [6,7].

Two persistent obstacles hinder progress. First, there is limited end-to-end evidence that connects precedent retrieval, controllable tileable generation, semantic segmentation and vectorization, and geometry-aware mapping into CAD environments [3,8]. Second, existing demonstrations often fragment these steps across tools and ad hoc scripts, which obstructs reproducibility and transfer to practice [1,4]. Together, these obstacles make the literature rich in attractive images but sparse in validated, fabrication-aware processes [5].

This study advances a reproducible workflow that links historically grounded precedent retrieval to controllable periodic generation, followed by semantic segmentation and vectorization, and concludes with geometry-aware mapping into interior elements such as walls, ceilings, partitions, floors, and furniture. Precedent retrieval is implemented through a ResNet-50 and Vision Transformer (ViT) embedding pipeline, while tileable generation uses a LoRA-tuned diffusion model with motif and region controls. The pipeline enforces curve closure, minimum feature widths, and mapping strategies that respect joints and service paths, making outputs legible at intended scales and suitable for fabrication.

Beyond studio convenience, this topic intersects broader needs: contemporary interiors are expected to celebrate Islamic heritage while meeting code, fabrication, and budget constraints; public clients increasingly request respectful reinterpretation over superficial styling; and scholarship urges methods that make cultural/ethical checks explicit rather than implicit. Together, these factors motivate a reproducible bridge from images to buildable elements that are sensitive to regional diversity and calligraphy.

Prior work demonstrates classification, symmetry detection, and parametric reconstruction, and recent studies explore diffusion-based synthesis; however, most accounts validate steps in isolation, stop at raster images, or rely on ad hoc scripts that hinder replication. We target this gap with a documented, fabrication-aware pipeline and domain metrics aligned to interior deployment.

The research aims to address the following questions:

• Can a dual backbone (ResNet 50 + ViT) descriptor improve motif- and element-aware retrieval for Islamic interiors compared with single backbones?
• Does tile-aware diffusion with LoRA adapters yield higher tileability and symmetry coherence than non-tile-aware sampling and rule only baselines?

A depth-of-integration perspective guides development and evaluation, distinguishing surface-level styling from meaningful incorporation of motifs within spatial, tectonic, and programmatic constraints. Emphasis is placed on the last-mile transition from 2D imagery to 3D elements, converting what is typically a manual and error-prone redraw into a consistent and auditable operation. This framing responds to documented workflow fragmentation in prior AI-centered accounts [9].

Domain-specific metrics and human evaluation support the claims. Symmetry coherence, seam and tileability error, spline closure and junction valence, UV distortion, and feature width compliance quantify geometric soundness and fabrication readiness. Blinded expert ratings assess perceived integration depth relative to strong parametric baselines, and time-on-task is measured to characterize workflow efficiency from prompt to fabrication-ready geometry.

Cultural safeguards are integral to the pipeline. Calligraphy is treated as a semantic artifact rather than a purely visual texture, and regional balance is audited to avoid over-representation of specific geographies or schools. Curatorial credit and takedown pathways are outlined to address provenance and rights. These measures make cultural and ethical constraints explicit and auditable rather than implicit.

The contribution is threefold: a license-audited dataset schema and retrieval classifier for common Islamic motif families and architectural elements; a controllable, tileable generation stage tuned to heritage pattern grammars; and an image-to-geometry toolchain with evaluation metrics and a practical rubric that together enable adoption in interior architecture and conservation contexts [7,9]. The remainder of the paper details the materials and methods, presents quantitative and expert-evaluation results, and discusses limitations and implications for practice.

In addition to LoRA fine-tuning, utilizing a fine-tuned text encoder model was considered. This is very useful as the text encoder is the part of the overall model that understands and interprets the input prompt, and training one on important keywords like the different styles, periods, and calligraphy can greatly enhance prompt understanding [10,11].

2. Literature Review

2.1. Islamic Geometric Patterns

Islamic architecture is distinguished by decorative triads (geometry, vegetal arabesque, and calligraphy), which are interrelated yet analytically separable [12,13]. Geometric work, in particular, mobilizes repetition, symmetry, and proportion to suggest unbounded extension across regional schools while maintaining shared mathematical principles [14,15]. Canonical classifications map patterns by symmetry groups, construction rules, and modules, enabling historical comparison across sites and eras [9,12].

Within this broad field, several motif families anchor both scholarship and design practice. Girih strapwork organizes stars and polygons into interlaced fields that can propagate across large surfaces; evidence suggests that fifteenth-century artisans achieved quasi-crystalline tilings that anticipate Penrose logics [16]. Muqarnas modules articulate transitions between plan and section and vary across regions and periods [14,17]. Mashrabiya lattices serve as privacy and light-modulating screens whose geometry expresses an aesthetic–functional synthesis in building skins [18]. Arabesque vegetal ornament, though curvilinear, often shares geometric logic with polygonal fields, producing visual unity at multiple scales [12,14]. This corpus offers not only an archive for preservation but also a structured grammar for construction that contemporary computation can formalize and extend [19].

2.2. AI in Heritage Design

Computational analysis of Islamic patterning began with symmetry detection and group classification, laying foundations for automated cataloging. Early models identify plane symmetry groups and rotational or reflectional features in decorative tiles, enabling systematic indexing from photographs [20,21,22]. As machine learning matured, supervised classifiers learned to distinguish motif subtypes (e.g., star orders) with increasing accuracy, demonstrating that data-driven features can capture categories long curated by art historians [9,23]. Such tools scale documentation, enable the retrieval of comparable precedents across collections, and support segmentation tasks that separate the ornament from the substrate for archival or replication [21].

Generative and restorative applications extend beyond recognition to reconstruction and synthesis. Restoration pipelines infer missing ornaments from partial inputs, combining shape descriptors with neural predictors to reconstruct damaged geometric fields with high internal consistency [24,25]. At the building scale, computer vision integrated with BIM converts photogrammetric captures into geometric models that join heritage documentation with parametric editing [8]. On the creative side, rule-based grammars and algorithmic composition prefigured learning-based generative models, producing new rosette or star variations that remain legible within historical families [26]. More recently, GANs and diffusion models have been explored for culturally specific ornament synthesis; early studies indicate that learned models can reproduce symmetry-coherent distributions while offering stylistic breadth beyond fixed grammars [27]. In interior workflows, diffusion guided by structural cues (e.g., ControlNet) has been used to propose pattern ideas later translated into precise parametric representations, illustrating a division of labor in which AI explores while parametric tools guarantee geometric correctness [5].

A persistent limitation in the literature is workflow fragmentation: classification, segmentation, and generation are often validated in isolation with generic computer vision metrics, leaving open how outputs become fabrication-ready surfaces or elements for real interiors. Furthermore, generic quality scores (e.g., black-box realism metrics) may not capture what matters for patterned ornament: symmetry order, seam continuity for tileability, vectorization tractability, or mapping distortion when applied to 3D surfaces. Emerging work that couples learned generation with rule-constrained translation suggests a way forward, but comprehensive, design-ready pipelines and domain-appropriate evaluations remain underreported [5,9,28].

2.3. Parametric Design and Visualization Tools

Parametric design environments such as Rhino 3D with Grasshopper encode pattern logic as editable procedures, reviving compass-and-straightedge methods with computational speed and precision. Researchers have reconstructed historical patterns by scripting base grids, star polygon intersections, and repetition schemes, enabling faithful revival and controlled variation from a single generative description [29]. Beyond reconstruction, parametric studies demonstrate morphing across families via continuous parameter changes, supporting explorations that remain within the recognizable space of Islamic ornament while traversing forms impossible to draft manually at scale [30,31]. Shape grammars formalize these operations as rule sets (replace, subdivide, interlace) that software can apply recursively, yielding authentic-looking results without hand-tracing [32,33].

Parametric adaptability proves crucial when motifs move from flat panels to curved or irregular surfaces. By re-parameterizing boundary conditions, designers fit complex tilings to domes, shells, or bespoke wall geometries, with the software recalculating element placement and intersections [34]. Visualization and rendering platforms, 3ds Max, Blender, and Unity, close the perception loop by simulating light, shadow, and materiality for patterned elements. Studies of “digital girih” emphasize how perforated panels modulate light in ways central to spatial experience, making interactive visualization a design instrument rather than a mere presentation tool [1]. Daylighting toolchains (e.g., DIVA, Honeybee) quantify glare and illuminance as parameters like opening size, pattern density, or panel thickness vary, supporting evidence-based adjustments to lattice designs [35]. This integration of generative geometry with performance feedback has informed contemporary kinetic façades that reinterpret mashrabiya principles at building scale, prototyped parametrically before fabrication and actuation. In such cases, the ornamental logic and environmental function are designed together rather than retrofitted, underscoring the value of a unified geometric substrate [4].

Recent work has begun to braid AI generation into these parametric pipelines. In one example, Stable Diffusion guided by ControlNet generated pattern proposals that were subsequently translated into exact Grasshopper definitions for installation in a community center interior, combining exploratory breadth with manufacturing precision [5]. This hybrid model aligns tools with their strengths: AI proposes stylistic and compositional variety at low marginal cost; parametric systems enforce symmetry, continuity, and fabrication constraints; visualization quantifies perception and, where relevant, environmental side effects. Yet, despite promising exemplars, the literature still lacks a full bench-tested bridge from AI outputs to vector geometry, UV mapping, extrusion, and module snapping within common CAD workflows, all evaluated with domain-specific metrics and human judgments.

2.4. Literature Synthesis and Research Gaps

Taken together, the literature establishes three pillars. First, Islamic patterns possess well-studied formal grammars and regional logics that can be analyzed and parameterized [12,14,15]. Second, AI methods can recognize, restore, and generate such patterns, but they rarely evaluate what interior designers most need: symmetry coherence, tileability, vectorization tractability, and mapping performance when designs are applied to 3D surfaces and materials [5,8,9,20,21,22,23,24,26]. Third, parametric environments can guarantee geometric rigor and fabrication readiness while supporting visualization and analysis, but they benefit more from a wider, culturally grounded design space than rule sets alone typically explore [1,4,29,30,31,32,33,34,35]. The opportunity, therefore, is to combine these strands in an end-to-end workflow that starts with heritage-aware retrieval, uses AI for controlled, tileable synthesis, and then translates outputs into CAD-ready geometry with explicit, domain-specific metrics and blinded human evaluation. That is the gap the present study aims to close.

2.5. Aesthetic Integration Framework

Islamic motifs in interiors are evaluated through an aesthetic lens (how geometry informs form, proportion, and composition) rather than through purely functional performance metrics [6]. The framework distinguishes Superficial, Semi-Integrated, and Deeply Integrated uses, reflecting the longstanding critique that ornament either sits as detachable decoration or fuses with design intent. It addresses concerns about tokenism, where motifs are pasted onto otherwise generic forms [7,13], and aligns with arguments for principled, contemporary reinterpretations of tradition [36]. Operational scoring anchors for these levels are defined in the Materials and Methods Section. The framework is applied across walls, ceilings, partitions, floors, and furniture (the primary canvases and object scales in interior work).

Walls: Superficial: printed or painted patterns or decals that do not affect wall form. Semi-Integrated: carved plaster, tiled panels, or applied lattices that add relief but do not set openings or structural rhythm. Deeply Integrated: wall geometry follows motif logic, for example, perforation fields or structural screens whose solids and voids are pattern-driven, so that removing the motif would alter the wall’s concept [6,37].

Ceilings: Superficial: graphic treatments on flat planes. Semi-Integrated: coffer layouts, suspended elements, or ornamental vaulting where the pattern shapes visible articulation but not the primary structure. Deeply Integrated: pattern generates the ceiling’s three-dimensional logic (for example, muqarnas, patterned grid shells), merging ornament and structure in a single rationale [7].

Partitions: Superficial: printed or etched motifs on solid dividers. Semi-Integrated: freestanding mashrabiya-like screens where the cutout pattern forms the element but remains replaceable. Deeply Integrated: partition lines and load-bearing screens derive from geometric tessellations that organize circulation and sightlines; the motif becomes a spatial framework rather than a removable object [36].

Floors: Superficial: rugs or vinyl prints applied as interchangeable finishes. Semi-Integrated: stone or ceramic inlays and zellige compositions that contribute depth and rhythm but do not determine room geometry. Deeply Integrated: floor grids and modules (radial or polygonal fields) govern furniture zoning, axis setting, and aperture alignment; pattern becomes a planning instrument, so form and circulation trace its logic [7].

Furniture: Superficial: patterned textiles, painted doors, or inlays that do not affect the object’s structure. Semi-Integrated: carved backs, ogee arched legs, or perforated doors that shape salient parts while the typology remains conventional. Deeply Integrated: the object’s structure is grown from a lattice or star polygon field; load paths, surfaces, and silhouette coincide with motif geometry, blurring ornament and engineering [8].

3. Materials and Methods

Research design: This study follows a design science methodology with mixed-methods experimental evaluation, combining objective, domain-specific metrics with blinded expert ratings to assess the build–evaluate artifact and its integration into CAD-ready geometry.

3.1. Workflow Overview

The workflow, as shown in Figure 1, proceeds in five stages: (1) precedent retrieval using a dual backbone embedding index to assemble motif- and element-matched references; (2) tile-aware generative sampling (LoRA-tuned diffusion) conditioned by motif/region tokens; (3) semantic segmentation to isolate ornament from substrate; (4) vectorization to closed B spline curves with minimum feature checks; and (5) CAD mapping that projects, tiles, extrudes, and perforates vectors into walls, ceilings, partitions, floors, and furniture with UV and feature width safeguards. Each stage emits auditable artifacts (image tiles, masks, SVG/DXF vectors, and 3ds Max scene files) and logs thresholds for re-producibility.

3.2. Dataset and Precedent Retrieval

A custom image dataset is assembled for Islamic patterned interiors with three label axes: motif family (geometric/girih, arabesque/floral, calligraphic), historical period (e.g., Umayyad, Ottoman), and architectural element (walls, ceilings, partitions, floors, furniture). Sources include public archives and architecture databases, complemented by curated web imagery [9,38,39]. Source images have a minimum long edge of 1024 px; for training/evaluation they are standardized to 512 × 512 px with color normalization and clutter reduction focused on pattern-bearing surfaces. Standard augmentations (brightness/contrast jitter, mild rotation) are applied; flips are used only when symmetry is preserved [40]. Splits are by site/project to avoid leakage.

Inclusion and exclusion criteria: Minimal occlusion, sufficient contrast for edge detection, and metadata for site, period, and region are required; images with ambiguous provenance or rights are excluded.

Taxonomy and annotation: Each image is labeled along two axes: (i) motif family (for example, star polygon, strapwork, arabesque, muqarnas, lattice) and (ii) architectural element (wall, ceiling, partition, floor, furniture). Optional tags capture sub-motifs and regional schools. Annotation quality is checked by dual-pass review with adjudication. Calligraphy is flagged as a separate semantic category to prevent misuse in generation and to route any text-bearing samples to restricted handling.

Dataset statistics and splits: The final dataset contains 7200 images (train 5040, val 1080, test 1080) with a minimum long edge of 1024 px. File formats are lossless or high-quality compressed.

Preprocessing and augmentation: Images are resized to 512 × 512 while preserving aspect ratio, center-cropped where appropriate, and normalized to the retrieval backbones. Augmentations include brightness and contrast jitter within conservative bounds, mild rotation, and flips only when pattern symmetry is preserved. For explicitly tileable samples, periodic padding is applied to maintain seam consistency during training.

Classification and retrieval: Motif recognition uses fine-tuned ResNet-50 and Vision Transformer (ViT) initializations from ImageNet [41,42,43]. Cross-entropy loss with early stopping and class weighting addresses imbalance. Penultimate layer embeddings are used for content-based retrieval (cosine similarity). ResNet 50 (2048 D) and ViT B/16 (768 D) features are L2 normalized, concatenated (2816 D), and projected to a 1024 D descriptor via a learned linear head. The ensemble option (late fusion of normalized logits/embeddings) improves precision without sacrificing interpretability.

Indexing and search: Descriptors are stored in a vector index using cosine similarity. Queries can be images, text prompts mapped to exemplar images, or hybrid queries that combine element constraints and motif hints.

Evaluation: Retrieval quality is reported mean average precision on the held-out test set, computed both globally and per element to surface failure modes. Confusion matrices are provided for the supervised head. Ablations compare single-backbone embeddings against the dual-backbone fusion, with and without class weighting. Failure analysis highlights confusions between closely related families and issues arising from dense arabesques or low-contrast relief.

Governance and licensing: Every image is tracked with provenance and license metadata. A takedown pathway is documented for rights holders. Sensitive or restricted materials are marked and excluded from public release. A data availability statement specifies which components are shared (schema, labels, example tiles) and which are restricted.

Reproducibility: Training uses fixed random seeds, documented hyperparameters, and specified hardware. All preprocessing scripts and the retrieval index builder are released with command-line flags and example configs so the index can be rebuilt from the shared portion of the dataset.

3.3. Generative Model

Stable Diffusion is adopted and specialized via LoRA fine-tuning; Stable Diffusion 1.5 is fine-tuned with LoRA adapters to enable tileable, motif/region-controlled synthesis. Periodic boundary conditions are used during training and sampling. Prompts are conditioned on motif family and region tokens; negative prompts discourage fractured strokes and broken symmetries. Raster tiles and high-resolution upscales are then exported for downstream vectorization [44].

Config: LoRA rank 16; learning rate 1e-4; steps 60,000; batch size 8; scheduler cosine with 5% warmup; tile size 512 × 512; CFG 7.0; sampler DPM++ 2M Karras; seed 42. Periodic padding and wrap-around sampling are enabled.

3.4. Image-to-Geometry Pipeline

Generated or retrieved motifs are converted into CAD-ready assets through three steps. (1) Segmentation: A Mask R-CNN model is fine-tuned to separate motif foregrounds from backgrounds across motif types [44]. (2) Vectorization: Edge detection followed by spline fitting yields clean, closed curves; minor smoothing reduces pixel jaggedness while preserving junctions critical to girih interlace. Edges are detected with Canny (low/high = 50/150); contours are fit to cubic B-splines with max deviation ≤ 1 px. Closure tolerance is 2 px; junctions below 35° are preserved. (3) Parametric mapping: In Autodesk 3ds Max V2025, custom scripts project, tile, extrude, and perforate splines on target surfaces, with controls for module size, minimum feature width, and UV layout. Units are millimeters. Minimum feature width is enforced at 4 mm; UVs are unwrapped with ABF (“Peel”). Scripts expose module snapping, collision checks, and auto-thickening when 4 mm is violated. The automation enforces fabrication-friendly topology and reduces manual redraw [5].

3.5. Application Across Interior Elements

For walls and partitions, perforation/extrusion depth and module snapping produce panels and screens aligned to the integration rubric. Ceilings use projection to planar coffers or to generated vault surfaces. Floors map tiles or inlay vectors onto planar meshes with grout-width parameters. Furniture workflows scale curves to part templates (backs, doors, legs) and check feature widths against material constraints. Each mapped asset is subsequently evaluated with seam and tileability error, symmetry coherence, vector closure, UV distortion, and feature width compliance as defined in Section 3.6.

3.6. Reproducibility and Controls

All models are trained with fixed seeds and documented hyperparameters; retrieval and generation scripts expose the same label ontology used for annotation. Outputs include both raster and vector assets, plus 3ds Max scene files with parameter states saved for auditability. Ethical safeguards for calligraphy (text-conditioned inputs and expert review) and regional balance are handled at data curation and prompt-conditioning stages, aligning the pipeline with the aesthetic integration aims articulated in the rubric [6,7]. Experiments ran on 1× NVIDIA RTX 3090 (24 GB VRAM), AMD Ryzen 9 5950X, 128 GB RAM, Ubuntu 22.04 LTS.

Versions: PyTorch 2.3.1, CUDA 12.1, Transformers 4.41.1, OpenCV 4.9.0, Python 3.10.

3.7. Objective Metrics

Acceptance thresholds: Unless otherwise noted, assets passing the following thresholds are deemed fabrication-ready: seam error EseamE_\mathrm{seam}Eseam < 0.02 (0–1 normalized); vector closure rate ≥ 0.98; atypical junctions (non 3/4 valence outside ±10° tolerance) ≤ 1% of nodes; median UV angle distortion ≤ 10° and area distortion ≤ 15%; feature width compliance FwF_wFw ≥ 0.995 at 4 mm minimum.

4. Results and Evaluation

This section evaluates the workflow end-to-end (retrieval, generation, 2D to 3D translation, and interior application) using metrics that reflect the patterned ornament’s specific demands and the aesthetic integration rubric. All training, validation, and test splits are made by site or project to avoid leakage. The annotated corpus includes motif labels (geometric or girih, arabesque or floral, calligraphic) and element labels (walls, ceilings, partitions, floors, furniture), with standardization and augmentation as in prior image pipelines [36,40]. Central tendencies with variance are reported where appropriate, and qualitative error analyses illuminate failure modes.

4.1. Retrieval and Classification

Motif recognition uses fine-tuned ResNet-50 and Vision Transformer (ViT) initializations from ImageNet [41,42,43]. Classification performance is summarized with accuracy, macro precision, recall, F1, and confusion matrices. Consistent with earlier Islamic pattern studies, the ensemble reliably separates geometric and arabesque categories while showing residual confusion between highly stylized calligraphic forms and vegetal curves [9]. Penultimate-layer embeddings are used for content-based image retrieval, with Recall@K and mAP reported; this supports rapid precedent discovery by motif and element. Qualitatively, retrieval returns regionally coherent sets (for example, zellige-dominant clusters for floor tiles), and failure cases often reflect mixed-motif scenes or occlusions, echoing segmentation challenges noted in architectural imagery [21].

4.2. Generation Quality and Domain-Specific Fidelity

A LoRA-tuned diffusion model is sampled with periodic boundary conditions to produce seamless textures and panel motifs conditioned on motif/region. FID and IS are reported for their comparability with prior work, with scores in the low-tens FID and IS ≈ 3–4 range, competitive with cultural GAN baselines while offering stronger control over tileability; this aligns with observations that learned models can respect symmetry distributions when properly conditioned [2,45,46,47,48,49]. Because generic realism metrics are insufficient for patterned correctness, additional domain metrics are applied: symmetry coherence, tileability and seam error, edge continuity and closure, and mapping-related measures. Figure 2 shows the AI vs. Parametric example.

Symmetry coherence: Frieze/wallpaper groups are predicted on generated samples and compared to test set distributions, complemented by radial power spectrum peaks at expected rotational orders (for example, 8, 10, 12-fold in girih). Outputs exhibit correct dominant orders with fewer spurious peaks under tile-aware settings, consistent with the constructive rules documented historically [12].

Tileability and seam error: For tileable textures, opposite-edge similarity and perceptual seam error are computed; periodic sampling substantially reduces edge discontinuities relative to non-periodic baselines, with residual artifacts concentrated in dense arabesques where long tendrils cross tile boundaries.

Edge continuity and closure: Vectorization readiness is estimated by the fraction of closed polygons and junction valence statistics. Geometric motifs achieve high closure rates; arabesque sets exhibit occasional micro-gaps at tight curvature changes.

4.3. Segmentation and Vectorization

Mask R-CNN is fine-tuned for motif–foreground segmentation. Evaluation of a hand-labeled subset reports mIoU and boundary F-measure. Error analysis shows under-segmentation where motif and background share material (for example, carved stone with shallow relief) and over-segmentation where calligraphy overlaps arabesques, consistent with prior façade and ornament segmentation. Vectorization quality is assessed with Chamfer distance between raster edges and splines, curvature variance smoothness, and topology preservation (Euler characteristic). Spline simplification reduces control points without violating minimum feature widths, a prerequisite for CNC/laser workflows [5,21,50].

4.4. Three-Dimensional Mapping and Fabrication Readiness

In Autodesk 3ds Max V2025, scripts project, tile, extrude, and perforate splines on walls, ceilings, partitions, floors, and furniture. UV distortion (area/angle) is logged on planar and curved targets; minimum feature width is enforced; self-intersections and thin ligatures are flagged. For walls and partitions, perforated panels meet width constraints except at acute star cusps; guidance templates increase local thickness while preserving legibility. Ceiling mappings show low distortion on planar coffers; muqarnas-like forms require local re-parameterization to limit angular distortion. By element: Floors emphasize joint continuity, where grout-width control and edge-snap produce coherent inlay paths, with most errors where star points converge near room edges. Furniture parts (backs, doors, legs) use templates that reject features below material-specific thresholds; ogee-arched legs derived from interlaced curves pass strength checks when extrusion depth and filet radii are increased modestly. Across categories, the pipeline produces fabrication-ready vectors and solids with limited manual retouching.

All fabrication checks (feature width, closure, UV distortion, self-intersection) were validated in the CAD/CAM environment; no physical prototypes were fabricated in this study. We therefore characterize outputs as fabrication-ready in simulation, and we outline planned physical validation in Limitations/Future Work.

4.5. Human Evaluation Aligned to the Rubric

A blind study is conducted with n = 18 domain experts (architects, artisans, historians; ≥5 years of experience) and n = 30 lay participants. Stimuli include (i) generated-and-mapped outputs from the pipeline, (ii) a parametric baseline crafted from canonical scripts without learned generation, and (iii) historical/contemporary precedents used with permission. Participants rate authenticity, aesthetic quality, and depth of integration per element on Likert scales derived from the rubric. Inter-rater reliability (ICC) is computed; paired comparisons use non-parametric tests with Holm corrections; effect sizes are reported. Expert raters favor tile-aware diffusion outputs over the parametric baseline in authenticity and perceived integration depth for walls, partitions, and floors; ceilings are closer, reflecting the baseline’s strength in geometric rigor for vaulting. Calligraphic scenes receive lower authenticity when the generator is unconstrained; ratings improve when text-conditioned inputs are used and vetted by experts, underscoring the need for semantic control [29,30,31,32,33,51].

Workflow Efficiency

Time from concept to fabrication-ready vectors/solids is measured for matched tasks. The AI-assisted flow reduces time-on-task relative to manual parametric redraws, with the largest savings for dense girih and arabesques on walls and partitions. NASA-TLX scores indicate lower perceived workload during segmentation/vectorization; effort shifts to prompt/parameter curation and review of fabrication checks, consistent with reports that AI expands the feasible search space while parametric tools guarantee constructability [4,5].

4.6. Bias and Data Governance

Regional coverage and motif balance are audited, with counts and performance by stratum. As seen elsewhere in vision datasets, skewed distributions correlate with uneven performance (for example, fewer West African or Chinese-Islamic exemplars depress retrieval accuracy and increase generator mode bias). Sources and licenses are documented; a takedown policy is provided; restricted materials requiring attribution are marked [52].

4.7. Limitations of Metrics

While symmetry, seam, vector, and mapping metrics align with the rubric’s aesthetic focus, they do not capture environmental performance (for example, daylight, glare) historically mediated by mashrabiya-like elements. These are reported contextually where relevant, but the rubric remains aesthetic by design. Moreover, generic FID/IS under-estimate ornamental correctness; emphasis is therefore placed on domain metrics and human judgments grounded in documented construction principles [12,37].

5. Discussion

The evaluation indicates that combining tile-aware diffusion with parametric mapping closes the frequent gap between compelling imagery and buildable geometry. Retrieval accelerates precedent study; generation supplies stylistically coherent, seamless motifs; segmentation and vectorization yield clean curves; scripts translate those curves into panels, screens, coffers, inlays, and furniture parts that satisfy basic fabrication constraints. Expert preferences for the outputs, particularly on walls, partitions, and floors, suggest that the pipeline moves motif use beyond surface-level ornament toward forms where pattern logic meaningfully shapes composition, the very shift urged by critiques of tokenism [6,7,36].

Individually, components (ResNet/ViT retrieval, LoRA tuned diffusion, Mask R CNN, CAD mapping) are known. The contribution lies in (1) tile-aware, motif/region-conditioned generation tuned for vector translation; (2) a geometry-first translation layer that enforces closure and minimum widths and logs UV distortion, producing CAD-ready vectors/solids rather than images; and (3) a domain rubric and metrics that score aesthetic integration (symmetry, seams, junction valence) instead of generic realism. To our knowledge, prior accounts validate these pieces in isolation; we bench test the bridge that carries raster outputs across the last mile into fabrication with cultural safeguards and reproducible scripts.

Classification and symmetry detection have matured to the point of robust indexing [9,20,21,22,23]. Parametric methods, in turn, guarantee geometric soundness and adaptability [29,30,31,32,33,34]. The contribution here is the bridge: domain-aware generation that respects symmetry and tileability, plus an automation layer that transforms 2D results into fabrication-ready assets across interior elements. This hybrid reflects the complementary roles identified in recent AI and parametric workflows [5]: AI explores; parametric tools enforce rules; visualization tests perception and, where applicable, environmental effects [1,35].

Human ratings confirm that authenticity in calligraphic scenes depends on semantic correctness. Treating calligraphy purely as ornament risks pseudo-script that can be culturally inappropriate. A text-conditioned pipeline (feeding verified Arabic text and stylizing it) mitigates this risk, and expert review remains essential [51]. Practically, this means calligraphy is the area where human-in-the-loop safeguards are non-negotiable, even as geometric and arabesque motifs can be more freely synthesized within documented rules [12].

The bias audit highlights how regional under-representation depresses model performance and narrows stylistic variety, echoing known dataset phenomena [52]. For practice, transparency on provenance and licenses is as critical as quantitative metrics. Designers need to know not just that an image is plausible, but that its source is legitimate and that its style is regionally appropriate to the commission. A takedown policy and attribution requirements help maintain trust, and the dataset can be steered toward balance over time via targeted collection.

The symmetry and seam metrics serve as pre-flight checks (see Section 3.5): if rotational orders and seam continuity fail, downstream vectorization and tiling will cost time to repair. Vector and mapping metrics (Chamfer distance, UV distortion, minimum feature width) directly anticipate fabrication issues. Conceived this way, the evaluation suite is not merely academic**; it is a design tool** that reduces iteration cost and helps designers argue for, or against, deeper integration moves with evidence rather than taste alone. For example, a floor inlay that yields low distortion and continuous joints provides a stronger case for pattern-led zoning than a superficially printed rug substitute.

Three technical fronts remain open. First, while tile-aware diffusion aligns outputs with CAD needs, additional symmetry priors could raise floor quality (for example, architectures that are equivariant to dihedral or cyclic groups to stabilize rotational orders) [53]. Second, calligraphy requires a closer pairing of language and style; text-conditioned workflows should be extended to more scripts and ligature behaviors [51]. Third, although the rubric intentionally scores aesthetic integration, many applications (especially partitions and ceilings) also carry environmental roles; integrating daylight simulations and ventilation proxies would clarify trade-offs where ornament becomes screening or structure [35]. At the building information level, coupling the vector/solid outputs with BIM schemas would streamline handoff to documentation and fabrication, following heritage reconstruction precedents [8].

The pipeline supports studios and offices that aim to avoid pastiche while working quickly. Retrieval grounds proposals in precedent; generation broadens the option set; mapping automates labor-intensive redraw. Used with the rubric, teams can stage integration: start with semi-integrated elements (for example, partitions), validate with metrics, and review, and then escalate to deeply integrated walls, floors, or ceilings where motif logic organizes space. This ladder provides a pragmatic path from visual theme to spatial structure, addressing critiques that contemporary work often stops at the surface [7].

Although the focus is Islamic ornament, the structure of the workflow (dataset with regional or motif labels, tile-aware generation, symmetry-aware metrics, and vector-to-CAD mapping) mirrors efforts in other heritage domains. The parametric and AI blend that enables controlled variation while preserving rule-based integrity is broadly applicable, provided analogous grammars and cultural safeguards exist [4,5].

The evaluations show that an AI-assisted, tile-aware, and CAD-conscious pipeline can help move Islamic motifs from detachable decor to form-giving devices in contemporary interiors. The strength of the approach lies as much in its checks and balances (symmetry, seams, vectors, mapping, human judgment) as in the generator itself. When paired with calligraphy safeguards and dataset governance, the system supports the kind of respectful innovation that scholarship has long urged: designs that converse with tradition through its underlying principles rather than its surface alone [7,12,14].

6. Limitations and Future Work

Physical validation: Where prototypes are absent, results are CAD/CAM simulated; we plan controlled fabrication studies across materials (wood, metal, plaster) to evaluate edge charring, kerf effects, and structural behavior of perforated panels.

Calligraphy semantics: Script requires text-conditioned pipelines and expert review; we restrict generative use to approved phrases and route sacred texts to manual workflows.

Regional bias: Under-represented regions mildly depress retrieval and generation performance; targeted curation and active learning acquisition will improve balance.

Symmetry priors: Incorporating dihedral/cyclic equivariance may further stabilize rotational orders and closure.

Transfer to other heritage contexts: The structure (dataset, tile-aware generation, vector/CAD mapping, integration metrics) generalizes to other patterned traditions with culturally specific safeguards.

7. Conclusions

This paper presents a design-ready workflow that reconnects Islamic architectural motifs with contemporary interiors through an integrated pipeline: a license-audited, annotated dataset; a retrieval module based on ResNet-50/ViT embeddings; a tile-aware LoRA-tuned diffusion generator for controlled, seamless motifs; and a scripted 2D to 3D mapping process (segmentation, vectorization, projection, extrusion) in Autodesk 3ds Max V2025. In contrast to fragmented approaches that stop at classification or image generation, the workflow yields fabrication-ready geometry for walls, ceilings, partitions, floors, and furniture, and evaluates results with domain-specific measures (symmetry coherence, tileability and seam error, vectorization fidelity, UV distortion) alongside a blinded expert and lay study aligned to a depth-of-integration rubric. The findings indicate that the approach supports movement from surface-level ornament to forms where motif logic meaningfully shapes composition, particularly for walls, partitions, and floors, while ceiling cases remain competitive with strong parametric baselines.

Methodologically, the work complements rule-based pattern design by providing data-driven variation and tileable synthesis that feed parametric rigor. It also resonates with emerging research on the AI capture of architectural style elements, extending that lens to culturally specific ornament with explicit controls for symmetry and tileability. Practically, the pipeline reduces time-on-task from concept to fabrication-ready assets, shifts effort from manual redraw to prompt and parameter curation, and embeds checks (minimum feature widths, mapping distortion, seam continuity) that anticipate fabrication issues before they propagate to production. Governance measures (source licensing, regional balance audit, takedown policy, calligraphy safeguards) make cultural and legal constraints explicit within the design flow.

Several fronts remain for future work. First, semantic calligraphy: Treating script as meaning rather than pure ornament requires text-conditioned pipelines so that generated strokes encode the correct words and phrases; expert review is essential where sacred texts are involved. Second, geometry-aware synthesis: Incorporating explicit constraints and priors into generation can further stabilize rotational orders, closure of interlaces, and edge continuity; here, symmetry-equivariant features and graph- or layout-aware generation offer promising directions. Third, designer-facing control: A lightweight interface that exposes motif or region conditioning, complexity, module size, and mapping parameters would broaden access for practitioners who are not ML specialists, while preserving the auditability and guardrails built into the pipeline. Fourth, BIM integration: Exporting parametric panels, screens, and inlays with metadata (module, minimum feature width, material) would streamline documentation and scheduling and aligns with prior heritage BIM efforts. Finally, while the emphasis is Islamic ornament, the workflow generalizes to other heritage domains with appropriate datasets and safeguards. Existing results on textiles and symbolic systems suggest that analogous pipelines could support region-specific pattern design and preservation elsewhere.

The contribution is twofold: a technical bridge that carries AI-generated motifs across the last mile into CAD/CAM, and an evaluative frame that privileges aesthetic integration over superficial application. By combining retrieval grounded in precedent, tile-aware synthesis, and parametric mapping with domain-specific metrics and cultural guardrails, the approach advances a respectful, reproducible path for interiors where Islamic motifs are not merely placed on surfaces but help organize space and form. The result supports designers seeking to innovate within tradition (expanding the option set while maintaining rigor and accountability in contemporary practice).

The workflow translates AI-generated motifs into fabrication-ready (in simulation) vectors/solids with explicit checks for symmetry, seams, closure, and width, and yields expert rated improvements over parametric baselines for several element types. We provide scripts, configs, and checklists to encourage reproducible adoption and identify physical prototyping and expanded regional coverage as immediate next steps.