Definition of the architectural style metric: An approach to quantitative analysis of design using language-image model

Abstract

This study proposes Architectural Style Metrics (ASM), a CLIP (Contrastive Language-Image Pre-training)-based methodology for quantification of architectural design’s visual characteristics. To address the subjectivity and inefficiency of traditional architectural design analysis methods, ASM quantifies four visual features―curvature, saturation, transparency, and symmetry―by their relative positions between opposing states, generating comprehensive metrics including means, standard deviations, and feature outliers. The effectiveness of the methodology was validated through single/multi-image quantification and clustering analysis. The quantification consistently reflected perceptual visual characteristics, while clustering―based on a quantitative database of approximately 9000 images constructed in this study―yielded meaningful groupings with an average silhouette score above 0.5. The study further explores potential applications through the demonstration of ASM-based evaluation approaches for systematic analysis of architectural designs, among which classification achieved an accuracy of 87.2%. ASM offers interpretable and objective results without additional training or strict image constraints, enabling broad applicability. These results demonstrate ASM’s potential as a consistent and scalable methodology for data-driven design analysis.

Graphical abstract

Keywords

Architectural style metrics / Quantitative analysis / Language-image model / Façade design / Visual features

1 Introduction

Architectural design style serves as a mode of expression for the aesthetic ideals and cultural narratives of specific periods and societies (Chan, 1992). Scholars study the change, formation, and history of style to trace relationships among works and understand their cultural and historical contexts (Schapiro, 1961). Such research typically begins with visual observation of as building elements―form, spatial composition, and material use―which collectively define the stylistic composition of a work (Bafna, 2008; Krause, 2001). Then these stylistic elements are interpreted through lenses shaped by individual perception and disciplinary perspectives (Frampton, 1995; Giedion, 2009), which naturally leads to diverse understandings and interpretations even of the same work. While such observation- and interpretation-based qualitative studies enrich stylistic discourse, their selective emphasis on features and evaluative criteria often makes their findings difficult to synthesize or compare across different works, limiting their reuse in broader comparative analyses (Jahanian, 2016; Nasar, 1994; Rho et al., 2020).

Against this backdrop, the escalating complexity and variety of contemporary architectural production underscore the value of objective, quantifiable analysis methods (Groat and Wang, 2013; Fross et al., 2015; Shin and Lee, 2019). Previous efforts to classify architectural styles quantitatively have demonstrated potential (Demir et al., 2021; Yoshimura et al., 2019). However, the techniques employed often require extensive task-specific training, which makes them heavily dependent on dataset-specific conditions and limits their broader applicability (Queirós et al., 2017; Lee and Han, 2023). Recent advances in computer vision and multimodal AI have begun to mitigate these constraints by moving beyond single-modal models. In this context, language-image models leverage pre-trained knowledge and flexible frameworks, offering potential for more comprehensive analysis that can more effectively distinguish visual characteristics (Radford et al., 2021; Yang et al., 2023).

Building on these needs and opportunities, this study proposes Architectural Style Metrics (ASM), a set of quantitative indicators designed to capture the visual characteristics of architectural design style. The purpose of this study is to provide a consistent foundation for analyzing and comparing stylistic tendencies in architecture, enabling scalable and comparative analyses across diverse datasets and design contexts. To validate ASM, the research focuses on building exteriors as observed in visual materials, examining works of internationally renowned architects to demonstrate how the proposed metrics can support broader applications in style analysis and comparative architectural studies.

To this end, the research proceeds as follows: (1) theoretical investigation through literature review on existing AI and architectural style analysis methods (Section 2); (2) identification of key visual features from architectural theory and development of the ASM conceptual framework based on these features (Section 3); (3) development of a quantification algorithm and its implementation through the Language-Image model (Section 4); and (4) experimental validation through quantitative analysis and classification (Section 5). Section 6 concludes with a discussion of the framework’s potential and limitations.

2 Literature review

2.1 Conventional approaches to architectural style analysis

Architectural design style analysis has traditionally relied on descriptive and qualitative approaches, analyzing visual characteristics through buildings, photographs, drawings, and 3D models (Groat and Wang, 2013). These approaches interpret architectural features based on researchers’ expertise, contributing to the understanding of architecture’s complex nature (Venturi, 1966). They have been widely applied to distinguish and analyze various architectural styles.

Using this approach, numerous research studies and theoretical works (Calloway and Cromley, 1996; Fletcher and Polley, 2020; Hopkins, 2014) have analyzed and classified architectural styles by region, period, and movement. For example, Calloway and Cromley (1996) established style-distinction criteria by analyzing architectural elements such as doors, windows, walls, ceilings, floors, and lighting. They describe British Victorian style doors as “Front doors are panelled and sometimes arched in the gothic style. They were often green or woodgrained” (Calloway and Cromley, 1996, p. 236), illustrating how styles are identified through observable features including form, materials, color, and arrangement.

However, these human-driven approaches extend beyond fixed visual features, as they are shaped by perceptions formed through experience, cultural context, and aesthetic frames―sensory impressions and personal or collective notions of what feels appealing―which introduces subjectivity (Frampton, 1995; Norberg-Schulz, 1980). From a phenomenological perspective, architecture is perceived through changing sensory and environmental conditions, so that even the same building may be experienced differently depending on time, circumstance, or the observer (Rasmussen, 1964; Zumthor, 2006). From a semiotic perspective, architectural forms are understood within cultural and social codes, with their meanings interpreted differently according to the observer’s cultural background (Eco, 1986). Filtered through these experiential and cultural lenses, such an approach can yield divergent interpretations even of identical visual characteristics, making style analysis inherently subjective.

In addition to this perceptual and interpretive subjectivity, the conventional analysis approach is time-intensive and requires extensive expertise (Queirós et al., 2017). In architectural research, analytical criteria and metrics have often been devised for individual studies, tailored to specific objectives and datasets. While this enables each study to address its scope effectively, the absence of shared standards has resulted in fragmented and localized investigations, making it difficult to compare findings across studies or build cumulative knowledge bases (Chan, 1992; Groat and Wang, 2013; Grierson and Moultrie, 2011). These limitations indicate the need for standardized analytical methods capable of supporting large-scale, consistent style analysis, particularly as architectural production continues to expand in the digital age (Lee et al., 2024; Qian et al., 2023; Yang and Qian, 2025; Zhao et al., 2024).

2.2 Early AI-based approaches to quantitative style analysis

Quantitative analysis of architectural design styles transforms visual characteristics into measurable data, aiming to ensure objectivity, consistency, and comparability (Groat and Wang, 2013). Accordingly, quantitative methods have been explored in architectural style research to complement qualitative approaches. Early quantitative analysis originated with Birkhoff’s (1933) research, which attempted to quantify aesthetic value through mathematical formulas based on the ratio between order and complexity.

Subsequently, Chan (1994) defined architectural design style as a set of characteristics and explored their measurability. His research decomposed architectural visual characteristics into multiple variables and analyzed both the minimum number of features required to recognize a style and each feature’s influence on style recognition. The study extracted and statistically analyzed visual features from representative buildings of renowned architects to quantify patterns between styles. These early quantitative analyses were limited by manual data analysis processes and application scope.

The emergence of CNN models introduced new approaches to architectural style analysis. Yoshimura et al. (2019) developed methods for classifying interior spaces and building facade images by learning styles of 34 architects, while Kim and Lee (2020) created a spatial image classification system through learning interior design styles. Demir et al. (2021) detected visual design principles in artworks and architecture using deep convolutional neural networks.

However, CNN’s black-box nature makes identification of decision-making criteria in the classification process unclear (Lee and Han, 2023). While previous studies established style classification criteria by detailing specific features of target features (Kim and Lee, 2020; Yoshimura et al., 2019), their application in CNN models remains challenging as training labels are limited to comprehensive categories like style names. CNN-based analysis is also constrained to analyzing learned data (Krizhevsky et al., 2012), limiting its application to broader architectural design analysis. These limitations indicate the potential for alternative approaches in architectural design analysis using multimodal AI technologies.

2.3 Advancements of AI technologies in computer vision

AI technology in computer vision has evolved from single-modal to multimodal approaches. Early AI models focused on single-modal processing, where CNN architectures enabled visual pattern recognition through hierarchical feature extraction via convolutional layers (LeCun et al., 2015; Krizhevsky et al., 2012). While effective for basic image classification and object detection, these approaches were limited in understanding semantic relationships between different types of data.

The development of multimodal AI models marked a significant advancement in data processing capabilities (Ngiam et al., 2011). CLIP (Contrastive Language-Image Pretraining) represented a key innovation by mapping text and images into a shared embedding space using large-scale datasets (Radford et al., 2021), improving upon earlier models, such as DeViSE, VSE, and Show and Tell, that processed different data types separately (Frome et al., 2013; Karpathy and Fei-Fei, 2015).

Advances in language models and the Transformer architecture (Brown et al., 2020; Vaswani et al., 2017) further enhanced these capabilities, leading to sophisticated multimodal systems that combine visual and textual processing (DeepSeek-AI et al., 2025; Google et al., 2023; Wu et al., 2024; Yang et al., 2023). These developments provide new tools for analyzing relationships between visual characteristics and qualitative descriptions in architectural design. These advancements in multimodal AI technology offer new approaches to quantitative analysis (Jia et al., 2021; Ramesh et al., 2022), enabling systematic analysis of design aspects traditionally investigated through qualitative methods. However, research applying these techniques to quantitative style analysis remains scarce, as most current studies focus on captioning and image generation (Crowson et al., 2022; Hong et al., 2024; Yoo and Lee, 2025).

2.4 Directions for advancing architectural style analysis

Existing studies on architectural style analysis―ranging from human-driven approaches to CNN-based approaches―face persistent limitations. Human-driven approaches rely on subjective interpretation, while CNN-based methods, though numerical, function as opaque systems dependent on task-specific training sets and predefined labels. As a result, these approaches struggle to deliver stylistic analyses that are objectively measurable, interpretable, and scalable, particularly when applied to large or diverse architectural datasets.

Accordingly, there is a need for a framework that:

1) quantifies perceivable visual attributes as reproducible measures to enable objective and consistent analysis;

2) delivers results that are interpretable and can be transparently understood rather than treated as opaque outputs;

3) remains applicable across varied datasets, stylistic variations, and evolving contexts without being constrained by fixed categories or rigid data requirements.

The following sections present Architectural Style Metrics (ASM), a metric for quantifying stylistic attributes, derived through a multimodal quantification framework to enable scalable and interpretable analysis of architectural styles.

3 Conceptual modeling for quantification of architectural styles

3.1 Architectural style features for quantification

This section identifies design style components in architectural analysis and determines features for quantitative analysis. For this purpose, the research investigates existing visual analysis structures and establishes quantification criteria, focusing on design styles observable in visual materials.

\begin{array}{}\text{(1)}& DS=\left\{v{f}_i\right|v{f}_{i}isavisualfeature\}\mathrm{,}\end{array}

\begin{array}{}\text{(2)}& \begin{array}{ll}DS=\hfill & \{Curvature\mathrm{,}ColorSaturation\mathrm{,}Transparency\mathrm{,}\hfill \\ \hfill & Symmetry\text{,}\mathrm{...}\}\mathrm{.}\hfill \end{array}\end{array}

Ackerman (1963), Schapiro (1961), and Smithies (1981) defined style as an ensemble of distinguishable features. Following this definition, design style (DS) comprises visually perceivable features (vf_i) such as form, color, materials (Eqs. (1) and (2)) where their combinations form distinctive characteristics. Unwin (2009) and Forty (2004) argued that in actual architectural analysis, these element combinations can systematically explain and understand buildings’ unique visual characteristics. This approach provides a useful framework for analyzing architectural complexity systematically (Calloway and Cromley, 1996). Based on this approach, this study develops a quantification method by decomposing architectural design style into measurable features.

Style features include “physical form, pattern, or specific distinguishable characteristics of design” (Chan, 1994). This study focuses on basic features found universally in architectural designs and expressible in degrees of difference, rather than complex design features like “Extended cantilever.” This allows application of uniform evaluation criteria across all designs and efficient analysis of various designs without limitation to specific styles. These features were identified from architectural design style analysis literature, as shown in Table 1 (Calloway and Cromley, 1996; Fletcher and Polley, 2020; Hopkins, 2014).

Chan (1994) suggested that a style may be defined when four or more visual traits recur consistently. Building on this premise, this study focuses on four primary features discussed in architectural literature―curvature, color saturation, transparency, and symmetry―as a demonstrative scope for validating the proposed quantification approach. While forming the basis of this initial implementation, this selection is not intended as a comprehensive definition of architectural style. These features are chosen because they (1) are widely referenced in descriptive analyses of style, (2) are immediately perceivable in visual observation, and (3) exhibit clear semantic polarity, making them well-suited for contrastive prompt-based quantification (Arnheim, 1974; Calloway and Cromley, 1996; Ching, 2007; Hopkins, 2014; Rasmussen, 1964). Other visual features―such as proportion, repetition, complexity, texture, and pattern―are excluded at this stage due to their sensitivity to image conditions and their tendency to span multiple styles without providing clear distinctions. The framework is inherently extensible, and additional or alternative features can be integrated in future applications once validated for definition and computational stability.

The four selected features can be further categorized into form-related and surface-related attributes, each playing a distinct role in stylistic perception. Form-related features include curvature and symmetry. Curvature refers to the degree to which lines and surfaces forming a building’s exterior form are curved or linear (Ching, 2007). Historically, curvature has marked key stylistic shifts, from the dynamic, flowing forms of Baroque and Rococo to the organic geometries of modernism, reflecting departures from rectilinear conventions (Ackerman, 1963; Jencks, 2002; Schapiro, 1961). Symmetry, denoting the degree of formal balance around a central axis (Wittkower, 1988), has long been tied to theories of order and proportion, particularly in Classical and Renaissance traditions, while continuing to serve as a reference point for visual stability despite the rise of asymmetry in contemporary design (Ackerman, 1963; Venturi, 1966).

Surface-related features comprise color saturation and transparency. Color saturation, measuring the purity and intensity of hues on a building’s exterior (McLachlan and McLachlan, 2014), often reflects material expression and visual identity, from the vivid ceramic cladding of post-modern façades to the muted, monochromatic palettes common in minimalist architecture (Porter and Mikellides, 2008). Transparency, referring to light transmission through building envelopes (Frampton, 1995), has played a defining role in modern and contemporary architecture, where extensive glazing and lightweight materials are used to reveal structure and visually dissolve the boundary between interior and exterior (Rowe and Slutzky, 1963). These four features have been repeatedly employed across periods and by different architects as recognizable visual signatures, making them effective axes for distinguishing stylistic tendencies in this study.

3.2 Conceptual modeling of Architectural Style Metrics

The Architectural Style Metric (ASM) proposed in this study is a quantitative indicator that represents the visual aspects of architectural design style. This metric aims to transform complex design visual features into objective and measurable forms and quantify building’s design style. By quantifying various visual features (vf_i) composing architectural design style independently, ASM provides a comprehensive framework for understanding design characteristics.

ASM is formally defined as a quantitative indicator of DS:

\begin{array}{}\text{(3)}& Score(DS)=\left\{vf{s}_i\right|vf{s}_{i}isavisualfeaturescore\}\mathrm{,}\end{array}

\begin{array}{}\text{(4)}& Metric(Score(DS))=ASM\mathrm{.}\end{array}

Each visual feature score (vfs_i) transforms qualitative characteristics into measurable values by determining their relative position between opposing extreme states. For example, curvature is measured on a spectrum between completely curvilinear and completely rectilinear states, with higher scores indicating closer alignment with the defined state (Eq. (5)). This approach allows for nuanced understanding of how different features contribute to the overall design style.

ASM is derived from vfs_i, expressing their statistical distribution (μ, σ) and identifying feature outliers (fo_±) that significantly differ from the average properties (Eq. (6)). The mean value (μ) indicates the overall state of all feature values, with higher means suggesting stronger alignment with defined states across different features. The standard deviation (σ) reveals how much these different features’ scores vary from this average, with larger values indicating greater variation between features within the design. fo_± then marks features that lie notably above or below the design’s average scores, allowing comparison of relative presence between different features.

\begin{array}{}\text{(6)}& ASM=[\begin{array}{c}\mu (vf{s}_i)\pm \sigma (vf{s}_i)|+f{o}_{+}\mathrm{,}-f{o}_{-}\end{array}]\mathrm{.}\end{array}

For example, when considering features like curvature and color saturation, a high mean indicates that features generally show high values―meaning the design tends toward defined states such as being both curvilinear and highly saturated. A high standard deviation would suggest significant differences between feature values, such as when a curvilinear design shows relatively low color saturation, or vice versa. In terms of feature outliers, if curvature is marked as fo₊, this reveals that the design is distinctly more curvilinear compared to its other characteristics, while if marked as fo_―, it indicates a notably rectilinear nature. When no feature outliers are present, it suggests that all features maintain similar levels of presence in the design. Through these measures, ASM summarizes the overall stylistic profile in a consistent format, regardless of the number of features considered.

While ASM provides a basis for objective evaluation as an indicator of a design’s visual properties, it does not assess design quality or appropriateness. The metric solely quantifies the extent to which certain visual characteristics are present; additional criteria are required to determine whether these characteristics are desirable within a specific context. For example, a high presence of curvature merely indicates that the design has more curved properties, not that it is superior design. ASM’s primary value lies in relative comparison between designs, where designs with significantly different ASM profiles likely possess very different visual characteristics.

3.3 Extension of Architectural Style Metrics for multiple image analysis

While ASM can be applied to single architectural images, style is rarely defined by a single instance. As Ackerman (1963) noted, multiple works from specific periods, regions, or individual designers often share recurring visual patterns within acceptable ranges, which together form the distinctive identity of a style. Capturing these patterns requires extending the ASM framework beyond isolated images to encompass image sets that represent a project or architect. This section formalizes the extension of ASM to multiple image scenarios and defines methods for aggregating and analyzing stylistic consistency and variation across related designs.

In this extended formulation, architectural style is represented not by a single image but by a set of K images (DS^multi) that collectively interpreted as a stylistic type. For each image, indexed by k∈ {1, 2, …, K}, the same visual features (vf_i) are quantified, resulting in a set of individual feature scores (vf{s}_i^{(k)}). These individual scores are averaged across all images to produce representative feature scores (vf{s}_i^{(k)}) for the style:

\begin{array}{}\text{(7)}& \begin{array}{l}Mean\left(Score\left(D{S}^{multi}\right)\right)=\{\overline{vf{s}_i^{(k)}}|\overline{vf{s}_i^{(k)}}istherepresentative\hfill \\ vf{s}_i^{(k)}acrossallD{S}^{(k)}\}\mathrm{.}\hfill \end{array}\end{array}

These representative feature scores capture the dominant characteristics of the style across multiple instances. These scores are then used to compute a multiple image-based ASM (ASM^multi), following the same statistical structure as the single-image case:

\begin{array}{}\text{(8)}& AS{M}^{multi}=\left[\mu \left(\overline{vf{s}_i^{(k)}}\right)\pm \sigma \left(\overline{vf{s}_i^{(k)}}\right)|+f{o}_{+}\mathrm{,}-f{o}_{-}\right]\mathrm{.}\end{array}

In this structure, μ denotes the overall stylistic tendency, σ measures the internal contrast between features, and the feature outliers (fo₊, fo_―) highlight attributes that are especially emphasized or suppressed within the style.

Beyond the ASM^multi, the system also evaluates intra-style variability by measuring how much each visual feature fluctuates across the K images:

\begin{array}{}\text{(9)}& \begin{array}{l}Variability\left(D{S}^{multi}\right)=\left\{\sigma \right(vf{s}_i^{(k)}\left)\right|\sigma \left(vf{s}_i^{(k)}\right)\\ isthestandarddeviationofvf{s}_i^{(k)}acrossallD{S}^{(k)}\}\mathrm{.}\end{array}\end{array}

Low variability indicates consistent feature presence across all instances, while high variability reflects variation within the style, potentially due to experimental design approaches or stylistic evolution. Through these measures, ASM^multi provides a comprehensive framework for analyzing how architectural styles develop and sustain distinct characteristics across related works.

4 Development of Architectural Style Metrics

4.1 Algorithm structure of Architectural Style Metrics

The quantification framework for Architectural Style Metrics (ASM) proceeds through three main stages―input, quantification, and output―as shown in Fig. 1. Within the Quantification stage, the process is further divided into three computational steps: Score, Aggregate, and Metric.

The input consists of an architectural façade image (DS) and a set of textual prompts representing opposing visual states (vs_i, ~vs_i) for each feature (vf_i). These prompts remain fixed across all evaluations to ensure semantic consistency. Image inputs include photographs, rendering, or drawings of façades, with optional preprocessing (e.g., cropping, color correction) applied when necessary to preserve visual clarity.

In Score step, the inputs are processed by Language-Image model-specially, CLIP (Contrastive Language-Image Pretraining; Radford et al., 2021)―in order to produce feature scores (vfs_i). CLIP is selected for its (1) computationally reproducible output under identical input conditions, a property critical for ensuring consistent and unbiased quantification across large-scale datasets, (2) zero-shot capacity that allows for prompt-based analysis without fine-tuning, and (3) high computational efficiency and accessibility compared to heavier models such as BLIP (Li et al., 2022) or LLM-based models (OpenAI, 2023).

Technically, the image (DS) is embedded by vision transformer, and each prompt (vs_i, ~vs_i) by language transformer of CLIP in a shared vector space. The image vector (V_DS) is then compared with each prompt vector (V_vsi, V_~vsi) using cos ine similarity (Hessel et al., 2021) to compute alignment scores for opposing feature states (vss_i; ~vss_i):

\begin{array}{}\text{(10)}& vs{s}_i=\frac{V_{DS}\cdot V_{v{s}_i}}{\parallel V_{DS}\parallel \parallel V_{v{s}_i}\parallel }\mathrm{,}\sim vs{s}_i=\frac{V_{DS}\cdot V_{\sim v{s}_i}}{\parallel V_{DS}\parallel \parallel V_{\sim v{s}_i}\parallel }\mathrm{.}\end{array}

These scores are then normalized using the Softmax function to produce a single feature score (vfs_i), representing the relative presence of the feature as a probabilistic value scaled between 0 and 1:

\begin{array}{}\text{(11)}& vf{s}_i=\frac{e^{vs{s}_i}}{e^{vs{s}_i}+e^{\sim vs{s}_i}}\mathrm{.}\end{array}

This score (vfs_i) quantifies the degree to which the image expresses the defined visual state (vs_i) over its opposite (~vs_i). The Score step thus encompasses both the semantic embedding process and the conversion of raw similarities into normalized, comparable feature values suitable for statistical evaluation.

In the Aggregate step, the system handles multiple image inputs (K > 1). In this case, each image is processed individually in the Score step, resulting in K sets of feature scores. Then, in the Aggregate step, these feature scores are averaged per feature using the mean to derive representative feature scores (vf{s}_i^{(k)}) and the standard deviation to measure variability (Fig. 2). For single image cases (K = 1), this step is naturally omitted and variability remains undefined.

In the Metric step, representative feature scores are summarized into an ASM. The system first calculates the mean (μ) and standard deviation (σ) across all feature scores. At this step, features could be weighted differently to reflect their varying prominence in different architectural styles (Chan, 1994). However, as default, equal weights are assigned to maintain consistent interpretation. This approach provides consistent representation regardless of the number of features. Alternative methods―such as octal or hexadecimal packaging (Salomon and Motta, 2009)―may also achieve similar consistency, though they are less intuitive for interpreting design characteristics. During this step, features that fall outside the mean distribution range (μ ± σ) are also identified as feature outliers (fo₊, fo_―), and are marked with + or ― signs for intuitive understanding:

\begin{array}{}\text{(12)}& f{o}_{+}=\left\{v{f}_i\right|vf{s}_i>\mu (vf{s}_i)+\sigma (vf{s}_i)\};\end{array}

The resulting ASM and feature scores consistently yield identical values under the same conditions, enabling reliable quantitative comparison. These metrics can be utilized for systematic analysis of design visual characteristics, with graphical visualization enabling more intuitive evaluation.

4.2 Representation of Architectural Style Metrics

Architectural Style Metrics (ASM) and feature scores (vfs_i) can be visualized using any method that maps multiple design features to independent axes, with parallel coordinates and radar charts being prominent examples (Fig. 3). Parallel coordinates provide an effective method for visualizing multidimensional data on a 2D plane, allowing simultaneous observation of relationships between multiple variables. While radar charts offer intuitive pattern recognition through radial arrangement of features, they may become less readable when displaying numerous features. Both methods effectively visualize the distribution and relationships of design characteristics, though the appropriate choice depends on data characteristics and analysis objectives.

This study primarily employs parallel coordinates due to their flexibility in handling varying numbers of features. As shown in Fig. 3, this graph maps feature types to the horizontal axis and feature value scales to the vertical axis, with feature scores mapped and connected sequentially along their respective axes. The connected feature scores can be interpreted as the design’s unique “frequency,” where similar line patterns indicate similar design characteristics. ASM’s mean and standard deviation are represented by solid and dashed lines in different colors, while for multiple image-based ASM (ASM^multi), the variability of each feature within the DS^multi is also visualized, enabling intuitive understanding of style consistency.

5 Demonstration

5.1 Quantification of architectural style in a single design

This section demonstrates the application of Architectural Style Metrics (ASM) to quantify design style in individual buildings. Building on the previously developed framework, a quantification system utilizing the ViT-L/32 CLIP model was implemented to derive feature scores and calculate ASM scores. Text prompts were carefully constructed by extracting relevant keywords from architectural theory (Burden, 1995; Ching, 2007) and optimized for semantic clarity. As summarized in Table 2, each visual feature―curvature, saturation, transparency, and symmetry―is defined through a pair of opposing state prompts. These prompts were applied consistently across all images to ensure uniform quantification.

The Example 1 (Fig. 4(a)) exhibits rectilinear forms, low saturation, high transparency, and overall symmetry while the image in Example 2 (Fig. 4(c)) displays curvilinear forms, low saturation, slightly lower transparency, and an asymmetric shape. The corresponding feature scores align closely with these perceptual attributes, demonstrating that ASM accurately captures distinct architectural characteristics at the individual image level. Moreover, images captured under nighttime conditions produced feature scores comparable to those of daytime images (Fig. 4), as the system contrasts visual content with textual meaning, allowing the architectural attributes to remain consistently interpreted even when lighting or environmental conditions change (Hentschel et al., 2022). This result indicates that the system can reliably interpret architectural features across varying environments.

To examine the generalizability, objectivity and consistency of this system, further validation was conducted using two sets of images totaling 110 samples. These included feature modification sets, where individual features of building images were systematically altered, and comparison sets of buildings photographed from similar angles to enable comprehensive analysis. Results are presented in Tables 3 and 4. The analysis revealed that while modified features showed significant value changes, unmodified features maintained stability within ±0.05 on the normalized scale. These findings confirm ASM’s utility as an objective metric for architectural style analysis across diverse visual conditions within defined parameters.

5.2 Quantitative style definition from multiple designs

This section examines whether consistent stylistic characteristics can be quantitatively defined by applying the Architectural Style Metrics (ASM) framework. For this, we computed feature scores from 9601 images across 453 projects by 30 architects using the proposed quantification framework. These scores were aggregated at the project and architect level, resulting in a structured database annotated with corresponding feature scores and ASM values (Fig. 5).

Images used in this section were collected through stratified sampling to ensure both stylistic diversity and contextual coverage. Thirty architects―mostly Pritzker laureates―were selected based on professional standing, stylistic diversity, and the availability of documented works. From architects’ official websites and curated online image repositories, at least ten projects per architect were collected, with each project represented by a minimum of ten images (Fig. 6). Each image captures frontal, lateral, or corner façades, documented from one- or two-point perspectives under varied seasonal and lighting conditions, ensuring a broad and representative coverage of visual contexts. After collection, all images underwent filtering and preprocessing: those with significant visual occlusion or excessive digital alterations―distorting the building’s original form or color―were excluded, while images with multiple buildings were cropped to isolate the primary structure. Finally, all curated images were manually labeled by corresponding project name for subsequent analysis. Statistically, this curated dataset maintains a balanced and near-normal distribution, with most architects represented by 12–18 projects and 250–400 images.

Project-level feature scores were computed by averaging values from façade images. Figure 7 presents image-level scores and the aggregated ASM for Frank Gehry’s Dr. Chau Chuk Wing Building. The project combines curved brick façades with planar curtain walls―contrasting sharply in form, saturation, and transparency. These visual contrasts are quantitatively captured: brick sections produce relatively higher saturation and curvature, while curtain walls yield high transparency and rectilinearity. And the overall metrics are shaped by dominant features, moderated by less frequent ones. This demonstrates how uniform or heterogeneous façade compositions influence the aggregated stylistic profile, supporting ASM’s ability to capture nuanced stylistic identities within a single project.

Architect-level scores were obtained by averaging feature values across each architect’s projects. As shown in Fig. 8, Barragán’s work consistently demonstrates rectilinear forms, high color saturation, low transparency, and moderate symmetry, reflected in stable score distributions. This consistency is particularly evident in the high saturation scores across most projects, exemplified by the vibrant colors in Gilardi House (Project 12) with only Calvario Chapel (Project 1) showing notably different features through its predominantly white palette. Indeed, the Barragan Foundation (1996), an institution dedicated to archiving and studying Luis Barragán’s work, notes that the “Construction differs from Barragán’s design” in reference to Calvario Chapel.

In contrast, SANAA’s work exhibits consistently low saturation scores while featuring varying degrees of curvature and high transparency through glass curtain walls and translucent double-wall systems. Their stylistic evolution is quantitatively captured in the ASM (Fig. 9), from the rectilinear New Contemporary Art Museum (Project 9) to the organically curved Grace Farms (Project 11). Notably, their acclaimed Serpentine Gallery Pavilion (Project 13) demonstrates ASM scores closely aligned with their overall style average. These observations show that ASM allows the identification of stylistic coherence or diversity across an architect’s work―revealing consistent tendencies in some and varied strategies in others.

The ASM framework reliably captures stylistic consistency and variation at both the project and architect levels. It shows how repeated design strategies produce coherent stylistic profiles, while diverse elements contribute to measurable variation. These results support ASM as a valid basis for quantitative style analysis and classification. The method can also be extended to broader groups―such as regional, historical, or cultural styles―by aggregating façade image-based feature scores. For instance, Gothic buildings may show high transparency from stained glass, while Baroque examples may exhibit stronger saturation due to ornamentation. This adaptability highlights ASM’s potential for scalable, consistent analysis across large and diverse architectural datasets. By enabling standardized style profiling, ASM provides a quantitative foundation for further analytical and design-oriented research, including comparative analysis, clustering, and classification.

5.3 Style classification through Architectural Style Metrics-based clustering

This section validates ASM’s potential for systematic and quantitative architectural style analysis through multi-scale cluster analysis. Using the previously established ASM database, clustering analyses were performed at two scales: detailed style classification within individual architects’ works (intra-architect) and style classification among different architects (inter-architect). The ASM data consist of normalized, continuous vectors that can be represented as coordinates in a multidimensional space. This structure allows clustering using Euclidean distance and summarizing each cluster through centroids, which aligns with the operation of K-means. Based on this, K-means clustering was applied for the analysis, and the optimal number of clusters (k) was determined through silhouette analysis prior to clustering.

Examining SANAA’s works as a case study of intra-architect analysis, silhouette analysis indicated three optimal clusters (Fig. 10(a)). K-means clustering results (Fig. 11) classified SANAA’s works into three distinct groups: relatively opaque and rectilinear facades (cluster 9), transparent and curvilinear facades (cluster 1), and transparent and rectilinear facades (cluster 2). This classification aligns with the study by Martinez and Ramos (2022), which similarly identified distinct groupings based on the combination of transparency and curvature, quantitatively demonstrating these as key features of their architectural expression.

For the inter-architect analysis of thirty architects’ styles, silhouette analysis yielded nine optimal clusters (Fig. 10(b)). K-means clustering results (Fig. 12) grouped architects with visually similar characteristics together (Curtis, 1996; Trachtenberg and Hyman, 1986), such as minimalist designers Richard Meier and SANAA (cluster 0), organic form practitioners Zaha Hadid and Frank Gehry (cluster 4), and high-saturation rectilinear designers Luis Barragán and Aldo Rossi (cluster 6).

The clustering analysis results demonstrate ASM’s capability to consistently and objectively capture unique features of architectural styles. The system successfully quantified both style evolution within individual architects’ works and stylistic similarities among different architects, showing high consistency with qualitative classifications by architectural theorists. Simultaneously, silhouette scores computed from the actual clustering results (Fig. 13) further supported this validity. The average silhouette scores for the two analyses were 0.56 (intra-architect analysis) and 0.54 (inter-architect analysis), indicating that data within each cluster shared coherent visual characteristics, while data across clusters were distinct. These findings suggest ASM’s effectiveness as a tool for systematic and objective architectural style analysis, offering various applications from analyzing new architectural works to identifying relationships with existing architectural styles.

5.4 Future applications of Architectural Style Metrics

This section demonstrates the potential of Architectural Style Metrics (ASM) as a tool for quantitative design evaluation. Two primary applications are explored: style classification through database comparison and comparative analysis of multiple designs against specified criteria.

The first application enables style classification by comparing an input image’s ASM with a pre-established ASM database. The system identifies the most similar style by measuring feature similarity scores, with additional consideration of whether the input image’s ASM falls within the style’s variability range. Figure 14 illustrates result of the evaluation through parallel coordinates, where the blue line represents the input image’s ASM against the gray lines showing the ASM and consistency of the most similar architect’s style.

The second application enables quantitative ranking of multiple architectural images according to specified visual criteria. The system derives ASM scores simultaneously for all images and ranks them based on their similarity to either a database-sourced ASM or user-defined criteria. Results are presented through parallel coordinates visualization, where gray lines represent the reference ASM’s feature scores and variability.

Validation was conducted using 10 test sets, each containing 30 images (one per architect). The first application (style classification) required an average of 20 s to analyze each individual image, and the second application (ranking) completed analysis of entire 30-image sets in 1–2 min each. The system achieved 87.2% accuracy in the first application. For the second application, images were evaluated against specific style criteria, with the original architect’s works consistently achieving the highest rankings. The top-ranked results predominantly featured works from architects previously identified within similar style clusters, confirming alignment with prior architectural design clustering analyses (Fig. 15).

The demonstrated applications show ASM’s potential for quantitative architectural design analysis. By enabling systematic comparison with established styles and evaluation against specific criteria, this framework provides objective measures that complement traditional qualitative assessments. Building on this capability, ASM can support a wide range of academic and practical uses, including heritage assessments through stylistic profiling, regulatory and design reviews with quantifiable benchmarks, and urban planning that requires large-scale comparative analysis of building exteriors. These roles position ASM as a scalable and interpretable basis for architectural style studies, extending beyond the scope of the present demonstrations.

6 Discussions and conclusions

This research developed Architectural Style Metrics (ASM), a methodology for the quantitative analysis of visual characteristics in architectural design. Using multimodal AI, specifically CLIP, ASM quantifies visual features along defined axes and synthesizes them into metrics―means, standard deviations, and feature outliers―that consistently represent stylistic tendencies. By aggregating multiple images, ASM captures both individual design tendencies and broader stylistic patterns.

Building on these capabilities, the framework provides a computationally reproducible and scalable bases for style analysis. The accompanying database of 9601 images validates its use in profiling, clustering, and classification, achieving an average silhouette score above 0.5 and 87.2% classification accuracy. These results confirm ASM’s reliability as a foundation for downstream tasks such as analysis and classification.

In comparison to existing approaches, ASM addresses limitations in objectivity, scalability, and adaptability. Human-driven analyses remain rich in interpretive depth but require considerable time and subjective judgment (Queirós et al., 2017), while CNN-based methods (Demir et al., 2021) produce fast results but often depend on narrowly defined training data and categories. ASM, by contrast, requiring no task-specific retraining and operating through text prompts, generates consistent outputs across diverse datasets, serving as a supporting framework for automated and consistent analysis where conventional methods face constraints.

Nonetheless, several limitations define the framework’s current scope. First, it measures only four visual attributes―curvature, saturation, transparency, and symmetry―which simplifies the multifaceted nature of architectural style; expanding the feature set would enhance representational accuracy. Second, the current focus on individual architects should be extended to historical, cultural, and regional categories to assess generalizability. Third, standardized protocols for image acquisition and processing, together with large-scale quantified databases, are required to stabilize feature scores against contextual or photographic variations. Fourth, ASM relies on the CLIP model, which is sensitive to text prompts and may carry biases from its pre-training data, particularly for underrepresented forms, suggesting that ongoing adoption of more advanced or domain-specific models could help improve reliability. Finally, integration with architectural design tools and workflows is essential for practical deployment.

Beyond technical considerations, it is important to acknowledge that quantification can oversimplify the layered and context-dependent nature of architectural style. When styles are reduced to numerical traits, there is a risk of distorting cultural or historical contexts, introducing biases from AI models and unbalanced datasets, or encouraging misuse if stylistic scores are prioritized over interpretive analysis. Recognizing these risks highlights the need for ASM to serve as a complementary framework―enhancing scalability and consistency―while remaining integrated with qualitative and theoretical approaches to preserve interpretive depth and contextual understanding.

Despite these considerations, ASM offers substantial academic and practical potential. Its standardized metrics and database enable reproducible analyses of stylistic tendencies, supporting comparative studies, style libraries, and investigations into how design approaches evolve across architects and periods. In practice, ASM can provide consistent, data-driven references for design evaluation, regulatory review, and urban planning, while also informing heritage assessments and the curation of stylistic resources for professional use.

References

[1] Ackerman, J.S., 1963. Style. In: Ackerman, J.S., Carpenter, R. (Eds.), Art and Archaeology. Prentice-Hall, pp. 174–186.

[2] Arnheim, R., 1974. Art and Visual Perception: a Psychology of the Creative Eye. University of California Press.

[3] Bafna, S., 2008. How architectural drawings work―and what that implies for the role of representation in architecture. J. Architect. 13 (5), 535–564.

[4] Barragan Foundation, 1996. List of Works.

[5] Birkhoff, G., 1933. Aesthetic Measure. Harvard University Press.

[6] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Amodei, D., 2020. Language models are few-shot learners. Adv. Neural Inf. Process. Syst. 33, 1877–1901.

[7] Burden, E., 1995. Elements of Architectural Design: a Visual Resource. Van Nostrand Reinhold.

[8] Calloway, S., Cromley, E., 1996. The Elements of Style: an Practical Encyclopedia of Interior Architectural Details from 1485 to the Present, Revised ed. Simon & Schuster, Inc.

[9] Ching, F.D.K., 2007. Architecture: Form, Space, and Order, third ed. John Wiley & Sons.

[10] Chan, C.S., 1992. Exploring individual style in design. Environ. Plann. Plann. Des. 19 (5), 503–523.

[11] Chan, C.S., 1994. Operational definitions of style. Environ. Plann. Plann. Des. 21, 223–246.

[12] Crowson, K., Biderman, S., Kornis, D., Stander, D., Hallahan, E., Castricato, L., Raff, E., 2022. VQGAN-CLIP: open domain image generation and editing with natural language guidance. In: European Conference on Computer Vision. Springer Nature Switzerland, Cham, pp. 88–105.

[13] Curtis, W., 1996. Modern Architecture Since 1900, third ed.s. Phaidon Press.

[14] Deepseek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., et al., 2025. Deepseek-r1: incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.

[15] Demir, G., Çekmiş, A., Yesilkaynak, V., Unal, G., 2021. Detecting visual design principles in art and architecture through deep convolutional neural networks. Autom. ConStruct. 130, 103826.

[16] Eco, U., 1986. Function and sign: semiotics of architecture. In: The City and the Sign: an Introduction to Urban Semiotics. Columbia University Press, pp. 55–86.

[17] Fletcher, M., Polley, R., 2020. Architectural Styles: a Visual Guide. Princeton University Press.

[18] Forty, A., 2004. Words and Buildings: a Vocabulary of Modern Architecture. Thames & Hudson.

[19] Frampton, K., 1995. Studies in Tectonic Culture: the Poetics of Construction in Nineteenth and Twentieth Century Architecture. The MIT Press.

[20] Frome, A., Corrado, G.S., Shlens, J., Bengio, S., Dean, J., Ranzato, M.A., Mikolov, T., 2013. Devise: a deep visual-semantic embedding model. Adv. Neural Inf. Process. Syst. 26.

[21] Fross, K., Winnicka-Jasłowska, D., Gumińska, A., Masły, D., Sitek, M., 2015. Use of qualitative research in architectural design and evaluation of the built environment. Procedia Manuf. 3, 1625–1632.

[22] Giedion, S., 2009. Space, Time and Architecture: the Growth of a New Tradition. Harvard University Press.

[23] Google, Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., et al., 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.

[24] Grierson, D., Moultrie, C.M., 2011. Architectural Design Principles and Processes for Sustainability: towards a Typology of Sustainable Building Design. Design Principles and Practices, pp. 623–634.

[25] Groat, L., Wang, D., 2013. Architectural Research Methods, second ed. John Wiley & Sons.

[26] Hentschel, S., Kobs, K., Hotho, A., 2022. CLIP knows image aesthetics. Frontiers in Artificial Intelligence 5, 976235.

[27] Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y., 2021. CLIPScore: a reference-free evaluation metric for image captioning. ArXiv, abs/2104.08718.

[28] Hong, C., Gochuico, Y.A., Shin, J., Ha, C., Marquez, L., Yoo, Y., Lee, J.K., 2024. Improving architect-specific building image generation using reinforced data processing. October. In: 2024 IEEE MIT Undergraduate Research Technology Conference (URTC), pp. 1–5 (IEEE).

[29] Hopkins, O., 2014. Architectural Styles: a Visual Guide. Laurence Kin Publishing.

[30] Jahanian, A., 2016. Quantifying Aesthetics of Visual Design Applied to Automatic Design. Springer.

[31] Jencks, C., 2002. The New Paradigm in Architecture: the Language of Post-modernism, first ed.s. Yale University Press.

[32] Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., et al., 2021. Scaling up visual and vision-language representation learning with noisy text supervision. July. In: International Conference on Machine Learning. PMLR, pp. 4904–4916.

[33] Karpathy, A., Fei-Fei, L., 2015. Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3128–3137.

[34] Kim, J., Lee, J.K., 2020. Stochastic detection of interior design styles using a deep learning model for reference images. Appl. Sci. 10 (20), 7299.

[35] Krause, C., 2001. Our visual landscape: managing the landscape under special consideration of visual aspects. Landsc. Urban Plann. 54 (1–4), 239–254.

[36] Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. ImageNet classification with deep convolutional neural networks. Adv. Neural Inf. Process. Syst. 25, 1097–1105.

[37] LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436–444.

[38] Lee, J.K., Yoo, Y., Cha, S.H., 2024. Generative early architectural visualizations: incorporating architect’s style-trained models. J. Comput. Des. Eng. 11 (5), 40–59.

[39] Lee, S., Han, J., 2023. Analysis of architectural design style using CNN output layer values. J. Architect. Institute Korea 39 (3), 23–30.

[40] Li, J., Li, D., Xiong, C., Hoi, S., 2022. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. PMLR, pp. 12888–12900.

[41] Martinez, P., Ramos, F., 2022. Emerging patterns in the construction of Sejima’s, Sanaa’s and Nishizawa’s domestic spaces, from 1987 to 2010. Front. Architect. Res. 11 (3), 375–401.

[42] McLachlan, F., McLachlan, E., 2014. Colour and contingency: theory into practice. Archit. Theor. Rev. 19 (2), 243–258.

[43] Nasar, J., 1994. Urban design aesthetics: the evaluative qualities of building exteriors. Environ. Behav. 26 (3), 377–401.

[44] Ngiam, J., Khosla, A., Kim, M., Nam, J., Lee, H., Ng, A.Y., 2011. Multimodal deep learning. In: Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 689–696.

[45] Norberg-Schulz, C., 1980. Genius Loci. Towards a Phenomenology of Architecture. Rizzoli, New York.

[46] OpenAI, 2023. Gpt-4 Technical Report. arXiv preprint arXiv: 2303.08774.

[47] Porter, T., Mikellides, B., 2008. Colour for Architecture Today, second ed.s. Taylor & Francis.

[48] Qian, W., Yang, F., Mei, H., Li, H., 2023. Artificial intelligence-designer for high-rise building sketches with user preferences. Eng. Struct. 275, 115171.

[49] Queirós, A., Faria, D., Almeida, F., 2017. Strengths and limitations of qualitative and quantitative research methods. Eur. J. Educ. Stud. 3 (9), 369–389.

[50] Radford, A., Kim, J., Hallacy, C., Ramesh, A., Goh, A., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I., 2021. Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. PmLR, pp. 8748–8763.

[51] Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M., 2022. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1 (2), 3.

[52] Rasmussen, S.E., 1964. Experiencing Architecture. MIT Press.

[53] Rho, J., Park, M., Lee, H., 2020. Automated construction progress management using computer vision-based CNN model and BIM. Korean J. Constr. Eng. Manag. 21 (5), 11–19.

[54] Rowe, C., Slutzky, R., 1963. Transparency: literal and phenomenal. Perspecta 8, 45–54.

[55] Salomon, D., Motta, G., 2009. Handbook of Data Compression. Springer.

[56] Schapiro, M., 1961. Style. In: Philipson, M. (Ed.), Aesthetics Today. World Publishing Company, pp. 81–113.

[57] Shin, J., Lee, J., 2019. Indoor walkability index: BIM-enabled approach to quantifying building circulation. Autom. ConStruct. 106, 102845.

[58] Smithies, K.W., 1981. Principles of Design in Architecture. Van Nostrand Reinhold.

[59] Trachtenberg, I., Hyman, M., 1986. Architecture: from Prehistory to Post-modernism, first ed.s. H.N. Abrams.

[60] Unwin, S., 2009. Analysing Architecture, third ed.s. Routledge.

[61] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need. Adv. Neural Inf. Process. Syst. 30, 1–11.

[62] Venturi, R., 1966. Complexity and Contradiction in Architecture. Doubleday & Company.

[63] Wittkower, R., 1988. Architectural Principles in the Age of Humanism, fourth ed.s. Wiley.

[64] Wu, Y., Hu, X., Fu, Z., Zhou, S., Li, J., 2024. GPT-4o: visual perception performance of multimodal large language models in piglet activity understanding. arXiv preprint arXiv:2406.09781.

[65] Yang, F., Qian, W., 2025. Generative architectural design from textual prompts: enhancing high-rise building concepts for assisting architects. Appl. Sci. 15 (6), 3000.

[66] Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L., 2023. The dawn of lmms: preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (1), 1.

[67] Yoo, Y., Lee, J., 2025. Development of an automated building visualization tool based on a fine-tuned model using LLM-Reinforced data. J. Korea Institute Spatial Des. 20 (1), 115–126.

[68] Yoshimura, Y., Cai, B., Wang, Z., Ratti, C., 2019. Deep learning architect: classification for architectural design through the eye of artificial intelligence. In: Geertman, S., Zhan, Q., Allan, A., Pettit, C. (Eds.), Computational Urban Planning and Management for Smart Cities. Springer, pp. 249–265.

[69] Zhao, L., Song, D., Chen, W., Kang, Q., 2024. Coloring and fusing architectural sketches by combining a y-shaped generative adversarial network and a denoising diffusion implicit model. Comput. Aided Civ. Infrastruct. Eng. 39 (7), 1003–1018.

[70] Zumthor, P., 2006. Atmospheres: architectural environments. surrounding objects. In: Atmospheres. Birkhäuser.

RIGHTS & PERMISSIONS

2095-2635/2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.