Contemporary architectural aesthetic preferences based on popular AI images and models of the Civitai architect community

Abstract

Architects often struggle to understand the architectural aesthetic preferences of their clients. At the same time, attention has been given to how individuals in artificial intelligence (AI) architect communities express these preferences. This study explores contemporary architectural aesthetic trends by analyzing content from the AI drawing platform Civitai. A multiphase research approach analyzed a dataset of 2,000 popular architectural images and 10 architectural models. In the literature review phase, 13 high-quality studies were selected to identify key indicators for evaluating architectural aesthetics. The study phase involved rigorous reviews of data characteristics, resulting in two sets of statistical results. The drawing verification phase utilized three of the most popular architectural renderings and models to validate the findings. The study reveals that ultra-modern and futuristic architecture is emerging as a new style, accounting for 69.9% of the total usage of architectural preferences. AI community users articulate their architectural aesthetic preferences, prioritizing a building’s structure and external environment. Specifically, the “structure prompt” comprises 31.11%, the “structure tag” 23.37%, the “environment prompt” 20.83%, and the “environment tag” 24.5%. These results confirm that the surrounding environment plays a significant role in shaping architectural aesthetics. This study reveals emerging characteristics in contemporary architectural aesthetic preferences by AI-generated architect communities, providing valuable insights for architectural design and urban planning.

Introduction

Artificial intelligence-based applications in the Architecture, Engineering, and Construction (AEC) industry are increasing globally. There is a growing trend to utilize artificial intelligence (AI) to enhance performance across various industries¹. However, the industry has been relatively slow to adopt large language models². Most of these applications focus on the stages of design and implementation. It provides designers with the opportunity to incorporate AI into various stages of the construction and design process. It directly or indirectly enables decision-makers to make informed design decisions and improve their designs^3–5. The design phase is one of the most critical and important phases in the industry. Since 2015, AI-driven support for conceptualizing architectural designs has experienced steady growth⁶.

AI image generation technology has great prospects in the field of architectural design. It opens up new possibilities for exploring creativity and aesthetics. While most of the research emphasizes optimizing architectural designs⁷, their adoption is skyrocketing within the architectural community⁸. Architects can utilize AI-driven drawing tools and applications to enhance design iterations, thereby streamlining their workflow. They can optimize their work by providing various options and solutions for their design components^7,9. The application of AI models trained on user preferences can enhance architectural appearance, design, and visualization¹⁰. The number of models related to architectural design in the AI architect community is also increasing. It enables architects to create sketches and renderings that align with the user’s aesthetic preferences¹¹. This, in turn, will improve overall design efficiency and quality¹². However, it also has limitations. One major concern is that AI platforms can mostly replicate existing aesthetics rather than produce truly original designs. Since AI models like Stable Diffusion are trained on large datasets composed of internet-sourced architectural images, their outputs are inherently shaped by existing visual datasets. But research on architectural aesthetic preferences remains limited.

There remains a lack of research on contemporary architectural aesthetic preferences from the perspective of AI-generated drawings. This study fills the gap by analyzing popular images and models used in AI-generated drawings. By doing so, it offers a fresh perspective on contemporary architectural aesthetics, moving beyond traditional studies. Therefore, this study examines contemporary architectural aesthetic preferences by analyzing popular images and models generated and shared within the AI architect community, with a focus on their visual characteristics.

Architectural aesthetics and AI-based drawings

Recently, the AEC industry has been rapidly developing, driven by the widespread dissemination of artificial intelligence technology⁶. The field of architecture has seen a series of innovative practices and theoretical achievements related to architectural AI drawing. Many researchers have noted that artificial intelligence can significantly enhance the efficiency of architects’ workflows^13–19. This indicates that the application of artificial intelligence in the AEC industry is lowering the learning threshold and reducing the cost of communication and collaboration. Architectural AI drawings can help people understand and utilize architects’ architectural plans more efficiently⁸.Additionally, researchers are continually optimizing the artificial intelligence models themselves. A study investigated how natural language text is converted into visual architectural designs²⁰. Studies have also identified and extracted fundamental patterns, styles, and design principles in generative artificial intelligence models^21–23. A study proposed a novel hybrid algorithm that combines agent-based simulation with deep learning, which can help architects efficiently and automatically manage the architectural design process²⁴.

Architectural aesthetic preferences

Architectural aesthetic preferences refer to individual or collective inclinations toward specific styles, forms, materials, and design elements in buildings and spaces²⁵. Recent research on aesthetic preferences in architecture has been highly effective. Some scholars have argued that integrating and blending buildings with their environment should consider aesthetics^26–29. Studies on architectural aesthetic preferences typically follow two main approaches: one focusing on extrinsic aspects related to the building’s subject and the other on intrinsic aspects of the building as an object. This study examines explicitly the extrinsic elements of architectural aesthetics.

Several factors have been identified as contributing to an understanding of people’s preferences for building façades. Research has identified originality, complexity, clarity, and meaning as key conceptual attributes^28,30. However, without professional guidance, it is often difficult for the public to express their architectural aesthetic preferences for the physical attributes of high-rise buildings^31,32. Studies have examined the key factors that influence architects’ aesthetic preferences for building façades^30,33. These factors include design intent, the building’s functional role as a public space, and its integration into the built environment. Additionally, proportion, material, and color also significantly influence architectural aesthetic trends.

However, whether complexity and symmetry play a role in these preferences remains a matter of controversy. Some researchers believe that organized complexity is an essential factor. People tend to prefer buildings with fractal geometry, finding them more visually pleasing and attractive²⁹. Research from a psychological perspective has shown that both natural and artificial environments can shape individuals’ perceptions and judgments of architectural aesthetics^34–36. These environments also affect people’s behavioral tendencies when adopting and designing buildings.

Studies have also explored the potential of artificial intelligence in influencing architectural design preferences. Pena et al.^7,37 and Xu et al.³⁷ examined diverse aesthetic preferences through an architectural lens, highlighting the subjectivity of evaluating architectural aesthetics among users. The research further emphasizes the need for more personalized AI-driven design tools, making a wider range of preferences available.

AI-based architectural drawing

AI-based architectural drawings can be categorized into two main approaches, i.e., architectural creativity and architectural drawing. Architectural creativity encompasses imaginative and artistic representations of built environments, often exploring unconventional forms, surreal elements, or stylized aesthetics⁷. On the other hand, architectural drawings include more grounded, realistic renderings that resemble traditional architectural outputs such as floor plans, building façades, urban perspectives, and section views¹⁵.

AI-based drawing is a rapidly evolving field. AI-assisted drawing helps users by generating ideas, supporting evaluation, and optimizing the design process. Previous studies have examined the increasing popularity and applications of artificial intelligence in the construction industry. AlOmani et al.¹³ introduced an AI-driven method that can generate design ideas from various images and aid in the development of architectural layouts. Karimi et al.³⁸ developed an AI tool that can retrieve and display digital sketches of architectural elements relevant to the user. This, in turn, helps enhance the user’s creativity. Similarly, Li et al.¹⁶ demonstrated how AI can improve design efficiency by assisting architects in defining architectural intent and form. Milošević et al.³⁹ explored the role of AI in conceptual design. Architects can create digital landscapes, evaluate creativity, and optimize efficiency and innovation. Patterson et al. introduced Automated Drawing Assessment (AuDrA)⁴⁰ This AI-based tool utilizes deep learning to assess architectural creativity, thereby enhancing design innovation.

Research has also focused on the role of AI in generating architectural outputs. Zheng & Yuan¹⁸ studied how artificial neural networks learn architectural forms to create and analyze design features. Leach⁴¹ found that artificial intelligence is gradually being integrated into all stages of architectural design. These AI-based technologies provide architects/users with new tools and methods while ensuring that the final design decisions remain with the users. In addition, Rahbar et al.²⁴ proposed a hybrid approach combining deep learning and intelligent modeling to generate architectural floor plans automatically. This, in turn, enables architects to try new creative designs and find better and more efficient design solutions. Based on this, this study examines contemporary architectural aesthetic preferences through the use of AI-generated architectural drawings.

Methodology

A multi-stage research methodology was employed to investigate architectural aesthetic preferences within the AI architect community systematically. The research process consists of three key phases: literature review, visual content analysis, and drawing verification. First, a systematic literature review is conducted to identify high-quality studies and shortlist evaluation indicators. The visual content analysis phase involves both a strict and broad review to gather data from Civitai. Finally, in the drawing verification phase, architectural renderings are generated to analyze and identify features within AI drawings (Fig. 1).

Fig. 1.

Multiphase methodology adopted by this study (Authors’ own, 2025).

Phase 1: systematic review

A comprehensive literature review was conducted using Web of Science and Scopus databases. Additionally, AI-powered academic search engines, including Consensus and Semantic Scholar, were utilized to supplement the grey literature. These databases have large repositories of scientific research articles, making them a preferred choice among researchers^42,43.

Table 1 presents the keywords selected after review by two authors. The data were retrieved from the host institution of the first author between January and October 2024. Figure 2 illustrates the PRISMA framework used for the literature screening process. PRISMA ensures the shortlisting of relevant and high-quality publications. It is a systematic process and limits subjectivity.

Table 1.

Keyword query.

Group	Keyword query
Architectural aesthetic preferences	(“architectural” OR “architecture”) AND (“aesthetic preferences” OR “aesthetic choices” OR “design preferences” OR “design choices” OR “visual preferences” OR “visual choices” OR “style preferences” OR “style choices”)
Architectural AI painting	(“architectural” OR “architecture” OR “building design” OR “construction design”) AND (“ai” OR “artificial intelligence” OR “machine learning” OR “neural networks”) AND (“painting” OR “art” OR “artwork” OR “digital painting” OR “generative art”)

Fig. 2.

PRISMA framework for literature screening.

Identification

Two databases (Web of Science and Scopus/ScienceDirect) were used to identify the number of papers for the systematic review. Web of Science and Scopus were selected due to their rigorous indexing standards and multidisciplinary coverage⁴⁴. The search was limited to studies published between 2020 and 2025, as the field of AI-generated design and architectural aesthetics is rapidly evolving. Two main keyword groups were used: architectural aesthetic preferences and Architectural AI painting (Table 1). Keywords were selected iteratively based on preliminary scanning of titles and abstracts in the target field. After removing duplicates, 31 were further screened.

Screening

The 31 articles were further shortlisted by identifying high-quality papers. Similarly, some papers were also removed as they were not related to AI and drawing. The total number of articles was 16, which were then assessed for further eligibility. Three papers were then further eliminated due to their similarity to other papers.

Included

A total of 13 papers were finalized for further analysis. The 13 papers that met the inclusion criteria were focused on architectural drawings and the use of AI tools for generating paintings/images of buildings.

Phase 2: content review

After conducting a literature review on the latest advancements in architectural AI drawing and aesthetic preferences, key architectural aesthetic preference indicators from these studies are identified and categorized by functional type into evaluation criteria. Widely recognized images and models within the AI community were sourced from the Civitai platform to understand prevailing aesthetic trends. The research was classified into strict and rough reviews for in-depth analysis. Origin 2024 software was used to generate heat maps and compare architectural designs.

Strict review: This phase involves careful manual screening and analysis of data that meets the predefined selection criteria. Through this rigorous process, 97 high-quality, valid data points were identified.

Rough review: An automated approach was employed to supplement the rigorous review.

Data source: Civitai was selected as the sole data source for this study due to openness, architectural content availability, and user interaction features. It is one of the largest publicly accessible platforms focused on AI-generated imagery, supporting both the sharing of generative models and prompt-based renderings. Unlike other AI art platforms, it offers open access to image prompts, user-generated tags, download statistics, and engagement metrics, including likes and collections.

Data scraping and collection: Python was employed to scrape and analyze publicly available data from the Civitai platform. The process was implemented in a controlled environment using Jupyter Notebook. The script navigated through search result pages using the query terms “architecture” and “building,” sorted by “most reactions.” For each item, the script retrieved the post IDs, the total number of likes (“hearts”), tags, and prompts (metadata), as well as the image file URLs. The automated method enabled the collection of a much larger dataset, yielding 2,000 valid data points.

Data validation: These data points were obtained after manually excluding duplicate URLs and posts, entries with missing metadata, and non-architectural or irrelevant content (e.g., images focused on anime). Images with fewer than 10 total reactions were also excluded. For data validation, a random 10% sample of the final dataset was manually checked for relevance and tag consistency by two authors to ensure accuracy. The resultant dataset helped offer insights into the architectural aesthetic preferences of the users.

Data classification: The architectural images analyzed in this study were categorized into four typologies based on user-generated tags and prompts within the Civitai platform: realistic, fantastical, classical, and futuristic. For this study, “classical” refers specifically to images that incorporate Greco-Roman and Renaissance-inspired elements, as labeled and tagged in the dataset. “Traditional” is used in cases where users explicitly identified vernacular or regionally embedded forms (e.g., courtyard houses, colonial revival, or similar styles). “Modern” and “futuristic” categories capture designs with streamlined, contemporary forms and highly stylized or speculative elements, respectively. The study uses these categories as operational definitions based on platform metadata and community tagging. However, the study notes that these categories are not intended as exhaustive or universal representation of architectural styles.

Phase 3: drawing verification

The data obtained from previous phases was used for drawing verification. AI-generated drawings were used for the visual content analysis. Stable diffusion was selected as the AI drawing generator due to its popularity and ability to produce diverse, high-quality images. Additionally, Civitai was used to generate images that were aligned with the research data.

Writing prompt words for AI-generated drawings can be highly subjective. Therefore, the process was automated to minimize such subjectivity. Data was fed to ChatGPT to generate prompt texts automatically. The prompts were used to access a collection of architectural renderings that can help in understanding contemporary aesthetic preferences.

Results

Literature review stage

This section introduces the search results of the literature review on architectural aesthetic preferences and architectural AI painting, as well as the current development trends discussed and summarized in Table 2.

Table 2.

Latest relevant research findings.

S/N	References	Main findings
1	(Ghomeishi, 2021)28	Conceptual attributes are incorporated into the study of architectural aesthetics.
2	(Prieto & Oldenhave, 2021)33	Mismatched aesthetic preferences among city center managers and other stakeholders.
3	(Coburn et al., 2020)45	It was found that different aesthetic dimensions (coherence, fascination, and sense of belonging) activate distinct regions in the visual cortex of the brain, suggesting that the brain has specific neural encoding for these dimensions.
4	(Mao et al., 2020)30	The study found that people are more inclined to gather around historic buildings with distinctive appearances, such as architectural styles and decorative elements, which attract their attention.
5	(A. Saka et al., 2024)42	The study found that GPT models can be applied throughout the entire life cycle of construction projects, including design, construction, and operation.
6	(Joern Ploennigs & Markus Berger, 2023)15	The integration of artificial intelligence and art in architecture will bring about a redefinition of architectural design processes and design methods.
7	(AlOmani & El-Rayes, 2020)13	This study has automated architectural layout design, providing a demonstration for the intelligent development of the architectural design field.
8	(Patterson et al., 2023)8,40	The AuDrA platform has introduced significant technological innovations to the field of creativity assessment and is expected to foster the advancement of creativity research and practice.
9	(Beyan & Rossy, 2023)8	AI-generated architectural imagery can facilitate better communication and collaboration with clients, stakeholders, and the public, improving the understanding and acceptance of design proposals.
10	(Rahbar et al., 2022)24	This research presents a promising hybrid approach that leverages agent-based modeling and deep learning to enable the automated generation of 2D architectural layouts while preserving designer control and satisfying design constraints.
11	(Zheng & Yuan, 2021)18	When trained on design data collected from existing buildings, the neural network can learn and infer the geometric design features of architectural designs. This enables the method to be applied to data-driven analysis and to generate architectural forms based on learned patterns from real-world examples.
12	(Ploennigs & Berger, 2024)21	The article demonstrates how generative AI models can be trained on large datasets of historical architectural designs to identify and extract the underlying patterns, styles, and design principles that characterize different eras and cultural influences.

The aesthetic preferences of buildings have long been a significant research topic in architectural design. This study examines 13 publications related to AI-generated architectural drawings and aesthetic preferences. The literature review findings indicate that while AI-generated drawings are increasingly being adopted in architectural design, research on their use in evaluating architectural aesthetic preferences remains limited¹⁵,⁴⁶. The review revealed that a wide range of factors influences architectural aesthetic preferences. Due to this complexity, a classification framework proposed by Belaroussi et al.²⁶ was adopted. This framework categorizes influencing factors into two broad groups, i.e., intrinsic and extrinsic. Intrinsic factors relate to the architectural object itself. These include physical attributes such as materials, structure, color, and function. While extrinsic factors are based on the perceiver’s perspective. These encompass human-related elements such as age, profession, and environmental concerns. However, some attributes, such as color or material, may carry extrinsic symbolic meaning based on cultural associations or environmental context.

This study did not focus on extrinsic factors. Therefore, the classification of intrinsic factors was refined for a more detailed analysis. To achieve this, intrinsic aspects were further subdivided into five key evaluation indicators. These indicators formed the foundation of the research framework. This classification provides a structured approach to understanding architectural aesthetic preferences. A systematic evaluation of these design elements can help in understanding AI-based architectural trends.

Visual content analysis

Based on the architectural images on the Civitai platform, the study identified the most popular architectural images in the AI architect community (UTC + 3 time zone, October 4, 2024, 18:30). The study was divided into two groups: a strict review group and a rough review group. The search keywords were “architecture” and “building.” In the first group, 100 images were reviewed based on the criterion of “most reactions,” with each image required to have a minimum of 10 likes or hearts. In the second group, 1,000 images were retrieved for each keyword using the same “most reactions” criterion to determine popularity. These were arranged in descending order with respect to download counts. The “most reactions” option ensures that the most effective data was obtained, as other options can produce many search results that have never been seen or evaluated.

In the strict review group of the study, data were collected from the platform, and a thorough review was conducted, resulting in a total of 97 effective images. After manual review and categorization, four types of buildings were identified: realistic, fantastical, classical, and futuristic. The terms were manually reviewed, and irrelevant items, such as scenes featuring humans or robots as the main characters, were excluded. Duplicate items were also excluded, such as identical images that had been reposted by the image publisher, which had already received high ratings. The lower-rated images were also excluded. The collected images were then classified and analyzed. During the classification process, the preference type of each image was evaluated, and each image was assigned two values: reality, fantasy, classical, or modern. A single-color heat map was created to illustrate the architectural aesthetic preference types within the AI architect community (Fig. 3). The darker shades indicate a higher intensity of preference for specific styles. This visual representation helps highlight which aesthetic categories are most dominant in the AI architect community.

Fig. 3.

Heat map of architectural aesthetic preference types.

In Fig. 4, the 3D scatter plot displays the spread and clustering of aesthetic preference values assigned to each image. Most data points fall below a value of 30, indicating a large number of low-engagement entries. A few outliers with values over 200 show significantly higher user engagement, suggesting a concentration of popularity among a small set of images.

Fig. 4.

3D scatter plot of architectural aesthetic preference types.

In the rough review group of the study, the high-frequency words were analyzed in the “prompt” and “tag” search results. In the search results for “prompt,” irrelevant parts were excluded. Words that appeared more than 100 times were selected, excluding prepositions, numerals, and descriptions of image quality. When searching for “architecture,” a total of 151 high-frequency words in the architecture prompt were reviewed. When searching for “building,” a total of 117 high-frequency words in the building prompt were reviewed. In the search results for “tag,” the data results were ideal for “prompt,” and words that appeared more than 30 times were selected. When searching for “architecture,” a total of 62 high-frequency words in the architecture tag were reviewed. When searching for “building,” a total of 50 high-frequency words in the building tag were reviewed. Architectural aesthetic preference evaluation indicators were classification criteria for the high-frequency words in “tag” and “prompt.” Since the search keywords “architecture” and “building” are synonymous, there is a phenomenon of repeated use of high-frequency words. Therefore, only one item is retained. Tables 3 and 4 present the statistical results for high-frequency words and preference types in the evaluation indicators. Figure 5 presents a scatter plot of the most frequently occurring words found in user-submitted prompts. These words are grouped according to the five evaluation indicators: structure, material, environment, decoration, and style. Figure 6 shows the frequency of tag words added after image generation, also categorized by the evaluation indicators. Environment-related tags such as “sky,” “tree,” and “street” appear most frequently, indicating a strong user emphasis on external context. The scatter plot highlights the prevalence of visual themes users identify in the generated images. This reinforces the finding that environmental aesthetics have a significant influence on perceived architectural appeal. The analysis revealed slight semantic overlap among some high-frequency terms. Terms such as “sky,” “night sky,” and “starry sky” were categorized under the “Environment” category. However, this overlap did not significantly affect the overall ranking of the evaluation indicators.

Table 3.

Statistics of high-frequency words for Prompt.

Evaluation indicators	Prompt	Count	Percentage
Structure	railing, windows, sidewalk, tower, balcony, terrace, stairs, vertical, hills, castle, extrusion building, path, fountain, settlement, skyscraper, tall, spectacular, highly, grand, magnificent, narrow, depth, magnificent, square, spectacular.	24	31.11%
Material	metal, marble, color, granite, concrete, white, tile, wood, foliage material, brick, glass.	11	14.28%
Environment	trees, lighting, vehicle, neon, traffic, mist, moon, waterfall, colonies, cliff, spaceship, colonial, sunset, atmosphere, night, canyon, scenery, clouds.	18	23.37%
Decoration	texture, glass, intricate, luxurious, elegance, intricate, vibrant, contrast.	8	10.38%
Style	realistic, design, modern, dramatic, futuristic, fantasy, cyberpunk, avant-garde, expressionism, ancient, reality, fiction, abstract, soft, inspired, mystery.	16	20.77%

Table 4.

Statistics of high-frequency words for Tag.

Evaluation indicators	Tag	Count	Percentage
Structure	Ruins, stairs, lamppost, house, castle, pillar, door, window, road, skyscraper.	10	20.83%
Material	Monochrome.	1	2.08%
Environment	outdoors, sky, cloud, tree, city, ground vehicle, motor vehicle, car, cityscape, cloudy sky, car, night, mountain, street, grass, night sky, vehicle focus, plant, sunlight, sunset, moon, flower, rock, snow, sun, scenery, starry sky, bird, river, forest.	30	62.5%
Decoration	Reflection, lantern.	2	4.16%
Style	East Asian architecture, realistic and traditional media, fiction, and fantasy.	5	10.41%

Fig. 5.

Grouped scatter plot of ultra-high frequency prompt words (Prompt refers to user-input text prior to image generation).

Fig. 6.

Grouped scatter plot of ultra-high frequency tag words (Tag refers to user- or system-assigned label after image generation).

Furthermore, this study also conducted a statistical analysis of the popular models in the AI architect community. The search keyword was “architecture,” sorted by “relevancy,” and a total of 395 results were collected. Notably, from the third search result, models frequently appeared to be used for hand-drawn illustrations and interior design. This reflects a trend, but it is not related to the purpose of this study. Among the top ten popular models, four meet the review criteria. Among them, two models are different versions of the same popular model. This study selected the newer model version with a lower download volume for participation in the test. The total number of downloads for the four popular models is 75,500, accounting for 69.9% of the total downloads of the top ten models, which is equivalent to 10,800 times. Therefore, three generative artificial intelligence models were collected after manual review, as shown in Table 5.

Table 5.

AI model for generating architectural renderings.

S/N	Name	Download	Collection	Comments	Tipping	Likes
1	ArchitectureRealMix	39,100	515	20	240	2800
2	architecture_Exterior_SDlife_Chiasedamme_V6.0	6600	116	3	44	606
3	architecture_Urban_SDlife_Chiasedamme_V6.0	8300	158	3	0	601

Drawing verification

The purpose of using the current most popular AI architectural models in this study is to verify the data from the visual content analysis phase. The expected result of AI-generated renderings is to ensure rationality and objectivity. In addition, since some high-frequency words have opposite meanings, prompt words were classified into two paragraphs, and parameter testing was conducted in the step of generating drawing results. The test consisted of eight architectural images, and the results were reasonable (Table 6), with no unrecognizable features in the architectural images.

Table 6.

Architectural AI Drawing parameter test.

The most popular AI architectural models in the architect community were selected, and Stable Diffusion was used. Based on the classification completed for “prompt,” a set of architectural renderings was generated as output (Table 7). To reduce the bias in results caused by human intervention, the editing of the “prompt” for generating architectural renderings was completed by ChatGPT3.5 Turbo. ChatGPT 3.5 Turbo organized the specific content of the prompt into two paragraphs for the natural language prompts used to generate architectural renderings.

Table 7.

AI-generated architectural renderings.

Discussion

Futuristic architecture is becoming a new style

By analyzing the three most popular architectural rendering models in the AI architect community, avant-garde architectural styles (defined here as unconventional, futuristic, and digitally progressive styles) were found to be the most prevalent, with their download volumes significantly surpassing those of other models. This strong preference for futuristic designs highlights the emerging influence of AI drawing platforms in shaping aesthetic patterns and design preferences within their user communities. Avant-garde architecture, characterized by its innovative, experimental, and forward-thinking design, embodies contemporary architectural movements that challenge conventional forms and push the boundaries of design creativity (Hagan, 2020). However, the frequent appearance of futuristic forms on platforms like Civitai may not always indicate genuine innovation but rather reflect patterns that the algorithm has statistically learned to associate with user engagement. Such patterns could potentially limit the creative scope of digital architecture, reinforce dominant trends while marginalizing emerging or culturally diverse styles that are underrepresented in the training data.

Previous research, based on questionnaire surveys, suggests that people generally prefer traditional architectural styles, often favoring familiar, symmetrical, and historically rooted designs^47–49. However, this preference does not appear in the AI architect community. The disparity between stated preferences in traditional surveys and observed behavior in digital design communities suggests a potential shift in architectural aesthetics when individuals interact with AI-driven design tools.

Architectural models generated by the AI ​​ architect community reflect the public’s perception of architectural aesthetics. Digital models are not limited in terms of creativity. The built environment, on the other hand, is constrained by structural feasibility, regulatory policies, and economic factors. Although such AI-generated designs may never be fully realized as physical structures, they can still serve as valuable indicators of emerging architectural trends.

Search results in the AI ​​ architect community further highlight this preference for futuristic buildings that dominate the platform, many of which feature complex fractal geometric patterns. The widespread popularity of futuristic models indicates a strong preference among users for futuristic and complex architectural forms. These complex forms are closely related to computational design, parametric modeling, and organic architectural principles. This observation aligns with previous research findings^29,30, which indicate that people prefer buildings with “organized complexity.” This complexity is a balance between complex design and structural coherence. The prominent representation of futuristic forms in AI-generated architectural renderings suggests that digital media platforms may be influencing and accelerating the adoption of contemporary design principles. As AI tools become more widely used in the professional and amateur design communities, they may help to redefine architectural norms gradually.

Users of the AI ​​ architect community can express their own architectural aesthetic preferences

Previous studies have shown that non-expert users often struggle to articulate their aesthetic preferences in architecture. Many lack the technical vocabulary or formal training required to effectively describe their design ideas^27,28. However, this study did not observe the same phenomenon. This suggests a shift in how users utilize AI-driven design tools. In AI-driven design environments, users seem to be able to express their aesthetic preferences intuitively rather than being limited by their lack of expertise. AI-based tools are interactive and help users communicate their ideas effectively, eliminating the need for specialized expertise.

A key factor in this change is the utilization of user input within the AI architect community. In the search data, the term ‘tag’ refers to words that users autonomously add after generating architectural renderings. The tag function automatically categorizes aesthetic and stylistic elements using built-in tools that can help users refine and contextualize their designs. Thus, tags are added after the rendering process. Therefore, these tags directly reflect the architectural preferences of users and reveal patterns in how individuals perceive and classify different design features. Similarly, a prompt refers to the words that users submit to the computer program before it generates a rendering. Prompts act as the initial input that guides AI-generated designs. Since they represent users’ intent before any visual output is created, they can also serve as even more indicators of architectural preferences.

The study found a consistency between user-generated results (rendered images) and user-defined inputs (prompts and tags). This suggests that users can effectively communicate their architectural aesthetic preferences. This alignment demonstrates that AI-assisted design enables users to bridge the gap between abstract preferences and visual representations (see Tables 3 and 4). Furthermore, the reliability of these results, confirmed by evaluation indicators, suggests that AI-driven design processes may enhance architectural creativity. AI platforms enable a broader range of users to engage meaningfully with architectural design.

People pay attention to Building structures and the external environment

Architectural aesthetic preferences refer to the tendency of individuals to favor certain types of buildings over others. These preferences are often based on elements such as form, structure, materiality, and environmental integration²⁶. Various cultural, historical, psychological, and environmental factors influence these preferences in relation to the built environment.

During the visual content analysis phase of this study, user-generated content revealed two primary themes: architectural structure and environmental context. High-frequency words focused on detailed descriptions of building façades, functional structures, streetscapes, and green spaces. This emphasis suggests that users do not view buildings as isolated objects. Instead, they consider them to be components of a broader architectural and urban landscape. These findings are consistent with previous research on architectural aesthetic preferences^34,36. Past research has highlighted the significant impact of a building’s surroundings on how people perceive and evaluate it. This study further suggests that architectural preferences evolve as people’s awareness of environmental and urban design considerations continues to grow. Users appear to pay close attention to how built form interacts with urban form and contributes to its overall aesthetic impact. The advancement of large language models can further help in identifying better visualizations and outputs^50,51.

Uneven data distribution on architectural aesthetic preferences in the AI ​​ architect community

This study found that AI-generated architectural renderings tend to strike a balance between realism and fantasy in design, favoring classical styles over modern ones. These findings support previous research^29,52, which suggests that people generally prefer traditional architectural styles (classical, vernacular, or regionally grounded forms that are familiar and historically rooted). The appeal of classical architecture may stem from its historical significance and professional credentials. It may also be related to the psychological comfort it provides through familiarity and well-recognized, established design principles.

To verify these results, the seven highest-valued samples were manually reviewed. Interestingly, these samples are consistent with the overall trends in the broader dataset, strengthening the reliability of our preference classification. Although the data distribution is not entirely balanced, the consistency of the results suggests meaningful insights into architectural aesthetic preferences.

Conclusion

This study examined contemporary architectural aesthetic preferences by analyzing AI-generated images and models on the Civitai platform. Data was collected using a multi-stage approach (literature review, visual content analysis, and drawing verification phases) to identify and visually present user preferences. The study examined over 2,000 architectural renderings and identified emerging design trends. The results revealed a strong preference for futuristic architectural styles, with a consistent emphasis on structural and environmental features. High-frequency keywords reveal a strong focus on structural elements, building façades, functional layouts, street scenery, and green spaces. This shift indicates an evolving aesthetic preference among contemporary users, as reflected in both prompts and tags.

The research contributes to a deeper understanding of how digital platforms reflect and influence architectural aesthetics. It offers a new perspective by combining AI-generated visual data with human input to assess design preferences at scale. For architectural practice, the findings can guide designers in exploring future-oriented styles that align with emerging public interests. For policymakers and urban planners, the results suggest the growing importance of incorporating digitally informed aesthetic trends into design guidelines and planning frameworks.

Limitations and future research

Due to objective reasons related to internet user privacy, the study was unable to analyze people’s architectural aesthetic preferences from the perspective of extrinsic aspects within the AI architect community. Some attributes, such as color or material, may carry extrinsic symbolic meaning based on cultural associations or environmental context. Inter-rater reliability was not used in this study, which could influence the choice of keywords. The style categories identified in this study are based on user-generated tags and prompts within the AI community. These are oversimplified complex stylistic traditions into broad labels (e.g., “classical” vs. “modern”). While these operational categories are useful for detecting trends within the dataset, they remain limited compared to the nuanced understanding of architectural styles in scholarly discourse. Future work may benefit from a more fluid or multidimensional model of classification that reflects the interplay between formal and contextual influences on aesthetic preferences. Moreover, AI-generated content often reflects existing aesthetic patterns learned from training data, which may limit the originality of architectural styles analyzed. While most samples reflect lower preference values, a few high-value samples exceed the combined total of the remaining ones. Another limitation of the study is the inability to determine the specific intent behind each model download. Users may access models for various reasons, such as personal interest, professional design, academic experimentation, or casual browsing. In future research, researchers can pay more attention to people’s architectural aesthetic preferences based on users’ demographics and their intended use.

Author contributions

Author statement: Z.W. Proposed the research idea and designed the experimental plan; was responsible for data analysis and initial draft writing; provided funding support; conducted the experiment and collected data; participated in the data analysis and revision work of the paper.T.W Monitored and was responsible for project management throughout the entire process; provided funding support; participated in the final review and finalization of the manuscript.K.O.V. Provided technical support and data validation; supervised the project throughout and was responsible for project management; and participated in the final review and approval of the manuscript.J.H. Provided technical support and data validation; supervised the project throughout, and was responsible for project management; participated in the final review and approval of the manuscript.

Funding

This paper has been supported by the RUDN University Strategic Academic Leadership Program.

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.