Research on the Application Effectiveness of Generative AI in Design Projects from Data-Driven and Sustainable Perspectives

Abstract

Generative AI is bringing revolutionary changes to architectural design. From data-driven and sustainable perspectives, this study introduces scientific data analysis methods to explore the specific application scenarios and effectiveness of generative AI in the early, middle, and late stages of architectural project design, while also examining its potential value in the field of sustainability. The research first synthesizes industry viewpoints through online data analysis. Secondly, it selects three typical practical architectural projects of different scales and types in which the author participated in comparative testing, recording the time, operational processes, and outputs required for schemes generated by the “traditional creative workflow” vs. the “AI-assisted workflow” at various stages. A multi-dimensional evaluation is conducted combining subjective questionnaires and objective performance simulation data. This study finds that generative AI can significantly enhance design efficiency and scheme diversity and guide the construction of sustainability dimensions, but challenges exist in quality control and technology integration. This research will provide an empirical framework and data benchmarks for architectural practitioners, clarifying a new design path of “data-driven–human–machine collaboration–sustainable optimization”, which holds significant reference value for promoting the transformation of the construction industry towards high efficiency and low carbon.

1. Introduction

The desire for digital transformation in the current construction industry is increasingly growing, and the impact, challenges, and future prospects brought by artificial intelligence image generators have become the focus of current discussion. According to the “2024 Global Construction Industry Digitalization Report”, the professional digitalization penetration rate is only 35%, with long scheme iteration cycles and high labor costs; on the other hand, sustainable design demand is surging, while sustainability performance simulations in traditional design processes often lag behind the scheme finalization stage, leading to compressed space for low-carbon optimization [1]. The explosive development of generative AI provides technical possibilities for solving the above problems, with image generation models such as MidJourney and Stable Diffusion being typical representatives. Preliminary application explorations have already appeared in the field of architectural design, but related research mostly focuses on technical demonstrations. AI-generated design images lack systematic effectiveness verification based on actual projects and have not incorporated “sustainable design” as an evaluation dimension, making it difficult to support large-scale industry applications [2]. Against this background, this study takes generative AI such as Stable Diffusion and Inspiration Rendering as entry points to conduct empirical research from the dual perspectives of data-driven and sustainability, aiming to bridge the “theory-practice” gap in the architectural application of generative AI and provide the industry with implementable technical application paradigms [3].

1.1. Development History of Generative AI in Design

AIGC stands for Artificial Intelligence Generated Content. It involves the acquisition, extraction, and collection of information for re-creation, outputting content such as text, images, and voice. It is a form of content representation following UGC (User-Generated Content) and PGC (Professionally Generated Content), possessing immense driving force for innovation in the field of architectural design [4]. Starting in the 1950s, the Dartmouth Conference first proposed the concept of “Artificial Intelligence”, and during this period, computer graphics technology began to penetrate the architectural field [5]. From the 1970s to the 1980s, Mr. Nicholas Negroponte, founder of the MIT Media Lab, proposed in his book “The Architecture Machine” the idea of establishing predictive models for dialog, envisioning conversation with intelligent machines [6,7]. During the same period, CAD gradually became the industry standard in architecture, transitioning architectural design from manual drafting to electronic drawing. This technological innovation provided a carrier for subsequent AI algorithms. After the 1990s, AI entered the “statistical learning” stage, and the architectural field also developed from computer-aided drawing towards aided decision-making. Entering the 21st century, generative AI experienced a third breakthrough development, evolving from early algorithms like genetic algorithms and particle swarm optimization to the rise in deep learning and the optimization of AI performance. In the architectural field, this first saw the birth of Grasshopper, enabling parametric design and scheme linkage, allowing for the realization of complex curved forms, such as those by Zaha Hadid Architects. Later, the deep neural network models proposed by Google DeepMind and the optimization of generative algorithms provided technical support for building energy consumption simulation and optimized design forms [8]. From 2021 to the present, generative AI has technologically evolved from algorithmic generation to diffusion models. Diffusion models represented by Stable Diffusion and DALL-E 2 broke through the limitations of previous stages, achieving efficient bidirectional generation of “text-to-image” and “image-to-image”. They are continuously expanding towards higher-dimensional mathematical models, integrating more accurate judgment, creativity, and multiple perceptions, enabling efficient construction methods for the design industry (Figure 1 and Figure 2). In the current design process, the relationship between designers and artificial intelligence is becoming closer; computers are not just design assistants but collaborators in design, serving as an important driving force for innovation in production and construction [9]. In the near future, generative large model technology will expand towards higher-dimensional data formats (

Table 1. Prediction of generative large model technology and related potential modal progress.

Time	Before 2020	2020	2022	2023	2025	2030
Text	Spam detection; translation; basic Q&A, a	Basic copywriting; first draft, b	Longer form; draft deepening, b	Fine-tuning in vertical fields (scientific papers, etc.) to achieve better results, c	The quality of the final draft is higher than the human average, c	The quality of the final draft is higher than that of professional authors, c
Programming	Single-line auto-completion, a	Multi-line generation, b	Longer form; better accuracy, b	Adapt to more calculator languages; apply to more vertical fields, c	Text generation (software) product, c	Ultimate text generation (software) product, with quality higher than that of full-time developers, c
Image			Art, logos, photos, b	First draft (product design, architectural design, etc.), b	Final draft (product design, architectural design, etc.), c	The quality of the final draft is higher than that of professional artists, designers, and photographers, c

Note: Feasibility of large models (a, preliminary attempt; b, implementation imminent; c, far beyond the peak).

).

1.2. Research Questions

Existing research has two key gaps: Firstly, in the entire process of practical, implemented projects within architectural design institutes, the specific application modes of generative AI, typically represented by Stable Diffusion, are not clear enough, lacking a closed-loop demonstration of “Scenario–Operation–Effect”. Secondly, its effectiveness evaluation mostly relies on subjective experience, lacks quantitative analysis supported by data science, and is not linked to sustainable design objectives [10,11]. Based on this, this study focuses on the following core questions:

(1)

Through comparative analysis of common large generative AI models in the current industry, which large generative AI model is more suitable as the core application tool in the practice of architectural design projects [12]?

(2)

How can suitable application scenarios and standardized operational processes for generative AI in the early, middle, and late stages of architectural projects be constructed?

(3)

Can the actual effectiveness of this technology across multiple dimensions such as design efficiency, scheme diversity, and sustainability be quantitatively verified through data science methods (cluster analysis and statistical testing)?

1.3. Research Scope and Objectives

This study focuses on the “schematic design stage” of architectural design (corresponding to the project’s early to late stages), excluding construction drawing design and the construction phase; project types selected are office buildings (high-rise office towers), cultural buildings (municipal libraries), and industrial buildings (logistics industrial parks)—three typical project types, all being design projects in China that are either already completed or under construction, with a sample size of n = 3, ensuring the representativeness and transferability of the research conclusions; at the technical level, a framework is constructed based on Stable Diffusion v2.8.13. It integrates ControlNet (including canny and depth modules) with LORA architectural style models (e.g., custom models for “Modernist Architecture” and “Low-Carbon Building”). To maintain focus, this study deliberately excludes other generative AI tools like MidJourney and DALL-E [13]. Secondly, potential risks and technical limitations identified during the application process are recognized, and the path for generative AI to advance architectural development is discussed.

2. Materials and Methods

2.1. Definition of Traditional and AI-Assisted Design Processes

The “traditional design process” referred to in this study is a linear design model based on hand-drawn sketches, SketchUp pro 2025 modeling, and manual rendering with software like 3ds Max 2025. Its process resembles the “linear progression” design process model described by Archer (1969) [14,15]. Design steps advance in the sequence of “problem definition–scheme generation–evaluation–implementation”, with stages in the process connecting in a unidirectional, progressive manner. Combined with the actual operations of architectural design firms in the schematic design phase, the specific process is divided sequentially into the requirement analysis phase, scheme conception phase (early stage), form expression phase (mid-stage), deepening and optimization phase (mid-stage), and result output phase (late stage). The early scheme conception phase primarily uses hand-drawn conceptual diagrams to roughly design spatial organization and architectural form. The mid-stage employs relevant drawing software for assisted design, clarifying the relationship of architectural masses, spatial layout, modeling, and the texture of the building envelope. Subsequently, opinions from various disciplines like structure are coordinated for scheme adjustment. The late stage uses professional rendering software to produce architectural renderings, during which materials, lighting, shadows, etc., are adjusted. If unreasonable aspects of the design occur at a certain stage, it is necessary to return to the previous stage for revision, resulting in high rework costs and a longer design cycle.

The “AI-assisted design process” integrates generative AI tools like Stable Diffusion, forming a cyclical design paradigm of “rapid generation–quantitative evaluation–instant iteration”. In the early scheme conception stage, large generative AI models are selected to replace hand-drawn sketches, generating multiple sets of scheme concepts in bulk, from which several more suitable design images are selected. In the mid-stage form optimization phase, ControlNet is used for precise shape control, LoRA architectural style models determine the architectural presentation style, and performance simulation is front-loaded to avoid rework in later stages. In the late stage of architectural design, after the final architectural form is determined, modeling tools are used to refine the design, and AI real-time rendering plugins are utilized to output the final results (

Table 2. Comparison of two processes.

		Traditional Design Process			AI-Assisted Design Process
	Phase	Core Task	Tools/Methods	Outputs	Core Task	Tools/Methods	Outputs
Preliminary Phase	Requirement Analysis	Sort out functions, planning conditions, owner preferences, etc., to form a requirements list	Review materials, site surveys and owner interviews	Feasibility analysis report and site analysis report	Sort out functions, planning conditions, owner preferences, etc., to form a requirements list	Review materials, site surveys, owner interviews	Feasibility analysis report and site analysis report
	Concept Design	Hand-drawn design sketches to explore architectural forms, styles, spatial organization, etc., based on professional experience, with multiple rounds of modification and discussion	Drawing paper, marker hand-drawing, scheme discussion meetings	2–5 sets of hand-drawn concept sketches	Use generative AI’s text-to-image function, select appropriate prompts, multiple iterations, and generate various design style sketches	Text-to-image in Stable Diffusion	10–30 design images of different styles and types
Mid-Phase	Form Expression	Convert selected sketches into 3D models, clarify building mass relationships, facade textures, and spatial layout; modifications require constant model readjustment; this phase is time-consuming	SketchUp pro 2025, Rhino 8, and other 3D modeling software, AutoCAD 2020 (for plan drawing)	3D models, plan layout drawings, elevation sketches	Determine the general design style, use ControlNet, LoRA, and other large model plugins for precise control, front-load the energy simulation phase, use AI-related software for testing, adjust building form and plan layout at any time, provide multiple design drawings for owner comparison and selection, and draw corresponding plan sketches	Image-to-image in Stable Diffusion, EnergyPlus, XiaoKu AI, and other energy simulation-related software, AutoCAD 2020 (for plan drawing)	Generate a set of series renderings with consistent volume and diverse details, plan layout drawings, and energy consumption assessment
	Deepening and Optimization	Develop detailed design drawings based on structural feasibility, regulatory codes, and analysis of factors such as natural lighting	Use SketchUp pro 2025 to deepen the model and Excel for energy consumption estimation	Scheme model deepened drawings and energy consumption estimation report
Later Phase	Results Output	Deliver the finalized deepened model to professional rendering agencies, and use professional rendering software to produce final results, requiring multiple rounds of adjustments and detail modifications during the process, forming presentable reporting results files	Lumion 12/3ds Max 2025 (used by rendering companies) and PPT (reporting files)	Final renderings and reporting files	Further deepen the finalized deepened model or selected design drawings from the previous stage using AI, form detail-rich renderings, and produce reporting results files	Inspiration rendering plugins and PPT (reporting files)	Final renderings and reporting files

).

2.2. Case Selection Criteria

To ensure the study’s representativeness, comparability, and practical applicability, the case selection follows four core criteria: different types, complete data, stage matching, and controlled variables (

Table 3. List of alternative cases.

Project Category	Project Name	Selected	Reason
Commercial	A Commercial Complex Project in Hangzhou	No	Incomplete Data
Office	Jinhua Information Economy Industrial Park	Yes	Meets Selection Criteria
Renovation	Key Building Enhancement Design North of Jinhua Huhaitang Park	No	Phase Mismatch; Renovation Projects Have Multiple Design Constraints
Residential	Residential Plot Project North of Jinhua Jinchuang Green Valley	No	Phase Mismatch
Industrial	Chengdu E-Workshop Project	No	Located in Chengdu; Regional Differences Exist
Cultural	Jinhua New Library	Yes	Meets Selection Criteria
Healthcare	A Medical Center in Yunnan Province	No	Located in Yunnan; Regional Differences Exist and Data is Incomplete
Industrial	Yiwu Comprehensive Bonded Industrial Park	Yes	Meets Selection Criteria
Education	The Supporting Secondary School in Shanghai Lingang New Area	No	Phase Mismatch; Multiple Design Constraints and Regional Differences Exist
Office	Yiwu Cross-Border E-Commerce Industrial Park	No	Weak Phase Compatibility; Prolonged Design Cycle Hinders Experiment Progress

Note: As some construction projects in the alternative cases are still in the confidentiality phase, their relevant visual materials (including renderings and actual site photos) are temporarily excluded from the presentation scope of this paper.

and

Table 4. Basic information of the three project groups.

Project Category	Case Name	Location	Current Project Progress	Final Scheme Effect	Site Area	Height Limit	Sustainability
Office	Jinhua Information Economy Industrial Park	Jinhua City, Zhejiang Province, China	Under Construction		1.68 hectares	≤80 m	Yes
Cultural	Jinhua Library	Jinhua City, Zhejiang Province, China	Completed		2.15 hectares	≤80 m	Yes
Industrial	Yiwu Comprehensive Bonded Industrial Park	Yiwu City, Zhejiang Province, China	Under Construction		45.84 hectares	≤35 m	Yes

).

(1)

In terms of type selection, empirical research needs to cover at least 2–3 different types of buildings to exclude interference from the particularity of a single type [16]. Among them, office buildings, cultural buildings, and industrial buildings are the three major typical project types in China’s current architectural design field. According to the “2024 China Architectural Design Industry Report”, these three project types account for 62% of the annual design volume, making them representative and ensuring that the research conclusions can be transferred to mainstream industry scenarios.

(2)

Regarding data completeness, the comparative experiment needs to obtain information such as modeling duration and discussion/modification frequency in the traditional design process, as well as generation time and prompt parameters for generative AI-assisted design. The lack of any item would prevent establishing a fair comparison baseline [17]. Among the considered types, commercial complex buildings were excluded due to missing iteration records of hand-drawn sketches.

(3)

Regarding stage matching, the core task in the initial stage of schematic design is ‘creative divergence—form exploration’, which is the efficiency bottleneck of the traditional process (accounting for 58–65% of the total schematic design cycle), providing a theoretical basis for this criterion in our study [18]. Among the potential cases, residential projects encountered by the author’s design institute had already entered the mid-design stage, and the building forms of residential types were relatively simple, making them unsuitable for complete application in the comparative experiment and unable to maximize the use of generative AI tools for scenario application; thus, residential building types were excluded.

(4)

For variable control, projects from the same region were selected as much as possible, adhering to the same set of local architectural design standards (e.g., “Zhejiang Province Green Building Design Standard” DB33/1092-2021 [19]), to exclude the impact of cross-regional standard differences in the design process.

2.3. Generative AI Model Selection and Adaptation

2.3.1. Analysis of Commonly Used Large Generative AI Models

In the schematic design process of design projects, it can be divided into early-stage task analysis, concept generation, and scheme exploration; the middle stage involves refining scheme details and comparing multiple schemes; the later stage mainly involves confirming the floor plan layout and presenting the final effect. Using suitable large AI models for different stages facilitates better expression in architectural design. The author selected four large generative AI models with currently high usage rates in architectural design institutes—MidJourney, Stable Diffusion, Modai Cloud AI, and SketchUp pro 2025’s built-in Inspiration Rendering plugin—and conducted analyses from dimensions such as core model principles, applicable stages, image generation comparison, and model training comparison. The following table analyzes the characteristics of several large AI models commonly used by architects (

Table 5. Comparative analysis of common AI large models.

Platform Name			Midjourney	Stable Diffusion	Modai Cloud AIi	SketchUp pro 2025 Inspiration Rendering (Plugin)
Core Principle Basis			Transformer-based deep learning generative model	Latent Diffusion Model (LDM)	Combines GAN and Transformer architecture	Ray tracing rendering engine + lightweight generative AI model
Applicable Stage			Early stage	Primarily mid-stage, also early and late stages	Early stage and mid-stage	Mid-stage and late stage
Image Generation Comparison	Usage Threshold	Hardware Requirements	Pure cloud deployment, no special configuration required	Local deployment preferred: requires mid-to-high-end GPU	Pure cloud deployment, no special configuration required	Relies on SketchUp pro 2025 runtime environment, moderate configuration requirements
		Ease of Use	Relatively low	Relatively high	Relatively low	Medium
	Feature Richness	Content Quality	Relatively high	High	Relatively high	High, excels in realistic style
		Architectural Industry Adaptability	Low	High	Relatively high	High
		Supports Text Prompts	Yes	Yes	Yes	Yes
		Strict Generation Control	No	Yes	No	Yes
	Extensibility		Low, only uses official templates	High, open-source, supports developing architecture-specific plugins	Medium, does not support third-party extensions	Medium, not extensible
Model Training Comparison		Usage Threshold	No custom training function	Medium difficulty	Medium difficulty	Requires official updates
		Training Feature Richness	None	High, supports Lora, Dreambooth, etc.	Low	None
		Training Effect (Architectural Scenes)	None	High	Medium	None

) [10,20].

Midjourney reduces the usage barrier through its open, free web-based platform and, by virtue of its distinguished artistic rendering performance, emerges as the optimal tool for inspiring ideation in the preliminary design phase. However, it lacks corresponding data training for the architectural industry, its generation results are relatively random, and it can only be used for providing design inspiration in the early stage [21]. Although Stable Diffusion requires local deployment and has a relatively high usage threshold, its powerful image-to-image and text-to-image capabilities and rich plugins demonstrate strong adaptability and control over architectural forms, supporting its use in all stages of architectural schematic design [22]. Modai Cloud AI is a specialized large AI model for the architectural industry, supporting the import of line drawings to quickly generate effects of specific architectural styles, suitable from early-stage conceptual design to mid-stage scheme presentation. However, this model cannot finely control form generation, and its system development needs improvement. The SketchUp pro 2025 Inspiration Rendering plugin is developed based on the mainstream modeling software in China’s architectural industry. This plugin enables rapid real-time rendering of 3D models, saving the need for model format conversion. However, it is weaker in detail adjustment and is more suitable for the visualization needs of mid-stage scheme presentations, team communication, and effect verification, efficiently connecting design and presentation phases.

Based on the data and analysis results in the table, Stable Diffusion and SketchUp pro 2025’s Inspiration Rendering plugin demonstrate superior applicability in the field of architectural design. In the author’s practice within actual architectural design projects, these two large models were also used as core application tools; subsequent project application practices and demonstration analyses will primarily focus on Stable Diffusion, with a small number of cases using SketchUp pro 2025’s Inspiration Rendering plugin.

2.3.2. Introduction to the Technical Principles of Stable Diffusion

(1)

Core Technical Architecture

Stable Diffusion is a model built on the Latent Diffusion Model for image generation. Its core advantage lies in transferring the spatial composition from high-dimensional pixel space to a low-dimensional latent space, significantly reducing computational costs while ensuring generation quality. The model architecture is mainly divided into three parts: Text Encoder, Diffusion Model, and Decoder [23,24]:

Text Encoder (CLIP Text Encoder): Converts natural language prompts, such as keywords like “high-rise office building, modernist style, good daylighting, shared space, low-carbon materials” into vector embeddings, providing targeted semantic guidance for image generation.

Diffusion Model: Inspired by thermodynamics, the diffusion model operates in a latent space through an iterative process of “noise addition and denoising” (typically 50–100 steps). It learns to generate feature data that conforms to textual semantics by repeatedly corrupting and reconstructing images (Figure 3). This process allows control over generation efficiency and precision by adjusting parameters such as the number of sampling steps and the CFG Scale (Classifier-Free Guidance scale), which regulates semantic alignment [25].

The Diffusion Model performs image processing tasks with results superior to GANs and does not require aligning and verifying distributions like VAEs. However, it has limitations in sampling steps and duration, requiring further improvement.

Decoder: Restores the feature map from the latent space into pixel-level architectural images, supporting output formats like PNG and JPG, with resolutions of up to 2048 × 2048 pixels.

(2)

Analysis of Adapted Functions for design projects

Stable Diffusion primarily uses three major functions for AI-assisted generative design: text-to-image, image-to-image, and control plugins. Text-to-Image: Uses precise prompts to directly generate architectural concept sketches. Key prompt content includes “Architecture + Style + Type + Main Morphological Features + Environmental Context + Sustainability Requirements + Other Supplementary Points”, for example, “Building, Modernism, High-rise Building, Office Building, Linear Form, Multi-block, Urban, Morning Light, Green Roof, Energy Efficiency Requirements” [22]. Compared to the pain point of “lack of inspiration” in traditional design, this function can quickly convert language into multiple images.

Image-to-Image: Generates design images based on existing 3D sketch models while retaining the original basic framework; uses the “Denoising Strength” function to optimize the scheme. For example, Denoising Strength = 0.2/0.3 basically keeps the original morphological features unchanged, with local detail adjustments and optimizations; Denoising Strength = 0.8/0.9 retains the general framework morphology, allowing for the exploration of diversified schemes [26].

Control Plugins: Mainly include ControlNet, LoRA, etc. (

Table 6. Models and scripts involved in the research and their uses.

Models/Scripts	Uses
LoRA Mode	Generate images with specific styles
X/Y/Z Plot	Generate comparison charts to find the best parameters
controlnet v1.1	Add image conditions to control pictures
Ultimate Plugin	Render images in blocks and add image details

). ControlNet is used to address the issue of excessive randomness in diffusion models by adding more image constraints, such as Depth, Lineart, etc., ensuring precise control over the image effect. LoRA (low-rank adaptation) is a sample training model, a low-rank adaptation method for large models, generating specific styles based on a small number of images (50–100), accurately improving the alignment between the scheme and project requirements.

The following shows the training process of the LoRA model (Figure 4).

2.4. Research Methodology

A mixed research method is adopted, integrating qualitative and quantitative analyses, specifically including four parts (

Table 7. Summary of research methods and tools.

Research Phase	Research Objective	Method	Tools
Preliminary Analysis	Understand current industry perspectives on AI applications in architecture	Web data crawling; semantic analysis	Python3.10; Large Model Prompt Engineering
Case Study	Compare the efficacy of traditional vs. AI-assisted workflows	A/B testing; case study	Stable Diffusion, AI-related plugins in SketchUp pro 2025, manual recording of
Data Analysis	Quantify scheme diversity and efficacy metrics	Cluster analysis; statistical testing	Scikit-learn, SPSS19, K-means, t-SNE, PCA
Multi-Dimensional Evaluation	Comprehensive evaluation of subjective and objective indicators	Questionnaire surveys; performance simulation	EnergyPlus8.1.0, Autodesk Ecotect Analysis 2013, SPSS19

):

2.4.1. Preliminary Analysis

Using Python 3.10, 80 articles strongly related to the theme of “AI and Architectural Design” were crawled from Zhihu (a Chinese knowledge-sharing platform), selected based on a click-through rate from high to low, to analyze and summarize viewpoints on the application effectiveness of generative AI in practical architectural projects. Each article was scored across eight dimensions [27]: Design Efficiency and Process Optimization, Quality and Innovation of Design Outcomes, Technology Integration and Acceptance, Compliance and Data Analysis, Cost and Resource Benefits, Multidisciplinary Collaboration and Communication, Technical Limitations and Practical Challenges, and Industry Impact and Professional Evolution. The score ranges from 0 to 10, where a higher score indicates stronger support for AI’s positive role in architectural design in that dimension, with 10 being “Most Supportive”, 5 being “Neutral”, and 0 being “Least Supportive”. If an article does not involve a dimension, it is assigned a value of −1. The scoring method is based on large models, using prompt engineering to set dimensions and scoring rules, and conducting semantic tendency analysis on the articles.

The reasons for choosing the Zhihu platform are as follows: The platform is an important community for Chinese architectural practitioners to share practical experience, and its content can effectively reflect the concerns and attitudes of the industry frontline; secondly, this study focuses on the “application effectiveness of generative AI in domestic Chinese architectural design practice” and needs to obtain localized data with “frontline practical experience” and “industry pain point feedback”. The user profile of the Zhihu platform happens to match this need, having a unique value for capturing the “practical voice” in China. However, the limitations of choosing this platform’s data objectively exist: the platform’s content lacks a peer-review mechanism, and some content sharing has “personal empiricism bias”. Therefore, for the crawled content, the final selection of valid data should, as much as possible, choose information supported by specific project cases; the platform’s lack of theoretical depth is also one of its limitations.

Regarding the selection of the eight evaluation dimensions, this is based on a review and extraction from the existing literature, and after consulting relevant experts, a systematic summary was conducted to cover, as comprehensively as possible, the impacts of AI on architecture. The author searched Scopus, Web of Science, and other relevant websites using keywords such as “AI design efficiency” and “architectural design evaluation”, selected articles published after 2021, screened 32 relevant empirical studies, and extracted multiple high-frequency dimension words such as “efficiency”, “innovation”, and “morphological rationality” (frequency ≥ 25 times). Secondly, five experts were invited for consultation and discussion to calibrate the necessity and independence of the listed dimensions. The experts include the chief architects from the author’s Class A design institute (two persons, each with over 15 years of architectural design experience and exposure to AI-assisted design), a university professor researching architectural design methodology and theory (one person), a project expert in architectural engineering management (one person), and a generative AI design tool development expert (one person). The selection of these experts covers key areas such as architectural practice, theoretical research, engineering management, and technology development, thereby ensuring completeness in professional depth and domain breadth.

2.4.2. Case Study

Three typical practical architectural projects of different scales and types in which the author participated were selected for comparative testing. The project types are high-rise office building, multi-story cultural public building, and low-rise industrial building. Using the A/B testing method, the projects were designed in parallel using the “traditional design institute creative process” and the “generative AI-assisted process”, respectively, setting up a comparative experiment and establishing a basic database.

By controlling variables, the differences in performance indicators between traditional and AI-assisted generative design were quantified to ensure the scientific nature and reproducibility of the experimental conclusions. The independent variables are the traditional design group (Group A) and the AI-assisted design group (Group B), ensuring the same number of designers per group and similar capabilities (all with 5 years of work experience, and each group is a team of three people). The dependent variables include quantitative indicators and qualitative indicators, which can be clearly divided into two categories (

Table 8. Quantitative evaluation criteria for traditional design and generative AI-assisted design.

Metric Category	Metric Name	Operational Definition	Traditional Process Measurement Method	AI Process Measurement Method
Time Efficiency	Preliminary Scheme Generation Cycle	Pure design time from “completion of requirements analysis” to the “first version of design sketches”, excluding time for requirement communication and presentation preparation.	Record the cumulative duration of all stages, including hand-drawing and modification.	Cumulative duration including prompt debugging, model generation, and design drawing selection.
	Mid-Phase Iteration Response Speed	The average time required to complete a single round of scheme modification and optimization, from form expression to deepening and optimization.	Calculate the cumulative duration of multiple typical modification rounds and take the average.	Calculate the cumulative duration of multiple typical modification rounds and take the average.
	Final Phase Generation Cycle	The time required to convert the optimized model into the final presentable design drawings.	The time spent by the rendering company on model optimization and processing.	The time taken to use generative AI-assisted software for rendering and generating the final design drawings.
	Decision Iterations	The number of modifications made during the design cycle, based on presentation rounds and internal discussions.	Record the number of complete modeling versions.	Count the number of modification rounds.
Resource Input	Input Cost	Labor cost is calculated based on daily salary; hardware energy consumption cost is the fee calculated from cumulative power usage; fees paid to rendering companies are calculated according to industry standards.	Calculate the fees paid to rendering companies and the labor cost of personnel reviewing the renderings.	Calculate labor costs and hardware energy consumption.

): (1) Quantitative indicators: Effective design time, labor cost, number of submitted scheme versions, and number of meeting rounds required to reach a decision. Data is collected through original records in the project process or tools. (2) Qualitative indicators: Scheme compliance, scheme innovation, and client acceptance. These indicators can be obtained through questionnaires, interviews, and communication with the client during presentations.

The test proceeded according to the “pre-test–formal experiment–data verification” process. The pre-test was mainly used to verify the feasibility of the experiment, arranging for designers to familiarize themselves with the AI-assisted design tools in advance to avoid experimental errors due to unfamiliarity with the tools [28]. Recent design projects were selected for pre-training to generate valid design schemes (as shown in Figure 5), and rules were clarified: if the AI-generated schemes require minor manual modifications, the manual modification time will also be counted in the total cycle.

The formal experiment involved issuing the same project brief to both Group A and Group B simultaneously, recording various indicators and collecting owner feedback. After submitting the final version, third-party scoring and evaluation were conducted according to pre-established scoring rules, while also collecting the client’s satisfaction score with the schemes. After the experiment was completed, an Excel spreadsheet or SPSS database was established, and data was entered according to “Project Number—Group—Various Performance Indicators” to ensure traceability of each group’s data.

Note: Since the selected projects are actual implemented projects, the start times, project milestones, and progression timelines of the three project groups are controlled by the owners. The total cycles are not uniform and cannot proceed simultaneously. This factor may affect the experimental results when comparing the three project groups horizontally.

(1)

In the Early Stage of Architectural Design Projects

Specific Operation: Mainly involves concept generation and style exploration. Therefore, the initial forms of the buildings were compared between two groups: the traditional design process and the generative AI group. The AI group selected Stable Diffusion, using the “text-to-image” function to transform abstract design requirements into concrete visual schemes. The traditional design group generated three schemes per project, while the AI-assisted design group generated and filtered out 30 schemes per project [29].

Regarding the Stable Diffusion operation process: Install the launcher for Stable Diffusion version 2.8.13. In accordance with the requirements of the task assignment, construct a prompt framework of “Basic Description + Style Constraints + Sustainability Requirements”; in the parameter settings, load the “depth” module in ControlNet to constrain the building scale, adjust the CFG Scale, Denoising Strength, number of generation steps, and resolution; generate 50 sets of concept sketches per project, retain 30 sets through manual preliminary screening (excluding schemes with morphological logic contradictions), and proceed to subsequent quantitative analysis (

Table 9. Comparison of parameters for architectural project schemes generated by Stable Diffusion.

Project Category	Case Name	Positive Prompt Description	Negative Prompt Description
Office	Jinhua Information Economy Industrial Park	(masterpiece, best quality, 8 k, detailed), modern office building, glass curtain wall, (daylighting: 1.3), rooftop solar panels, sustainable design, energy efficient LED lighting, natural ventilation, (employees walking: 1.1), surrounding trees, clear blue sky, isometric view, architectural rendering	(worst quality, low quality: 1.4), blurry, cartoon, anime, deformed, ugly, bad anatomy, extra limbs, (traditional style: 1.3), polluted sky, empty, no people
Cultural	Jinhua Library	(masterpiece, best quality, 8 k, detailed), modern public library, cultural building, large glass windows, (green roof: 1.3), rainwater collection system, (sunlit atrium: 1.2), eco-friendly wooden facade, public space, (reading citizens: 1.1), plaza, trees, low-rise building, top view, architectural visualization	(worst quality, low quality: 1.4), cartoon, sketch, overly ornate, classical style, dark interior, gloomy, crowded, chaotic, industrial park
Industrial	Yiwu Comprehensive Bonded Industrial Park	(masterpiece, best quality, 8 k, detailed), modern industrial factory, steel structure, large span, (energy-efficient skylights: 1.2), insulated walls, (solar panels on roof: 1.3), clean and tidy, ventilation, heat recovery, sustainable materials, minimalist design, daytime, aerial view	(worst quality, low quality: 1.4), dirty, rusty, broken windows, smoke, pollution, outdated, cramped, residential building, ornate decoration, text, human

) [17].

(2)

In the Mid- and Late Stages of Architectural Design Projects

Specific Operation: For the mid-stage of the project, based on the concept schemes filtered in the early stage, the traditional group uses SketchUp pro 2025 and other 3D modeling software, with Lumion rendering software for design; the AI group uses the “image-to-image” function in Stable Diffusion for deepening, generating schemes with different materials and facades. By adjusting parameters and keywords, different facade material variables are designed, ensuring minor adjustments and optimization of facade details while maintaining the overall conceptual direction. The generated renderings are compared to assist the design team and the owner in making quick decisions and simultaneously promoting the sustainable performance of the scheme. The specific operation process is as follows (Figure 6):

In the late stage of project design, the traditional group uses rendering companies to produce the final design drawings, while the AI group uses the Inspiration Rendering plugin, which generates more realistic effects, to optimize the details of the renderings, enhancing the realism of the image and the scene atmosphere. For the issue of refined energy consumption simulation, based on the mid-to-late stage deepened schemes and the national standards “Standard for building carbon emission calculation” (GB/T 51366-2019) [30], the EnergyPlus software is used to calculate the annual heating, cooling, lighting energy consumption, and total carbon emissions for each building scheme, clarifying the sustainability indicators of each scheme [31].

2.4.3. Data Analysis

Data analysis tools such as SPSS analysis software and the K-means algorithm in Sklearn are used to quantitatively evaluate the style diversity of the generated schemes, and professional chart analysis is generated to visualize the results.

In the scheme diversity analysis, the CLIP-ViL multimodal large model is selected for quantitative evaluation, automatically scoring the generated schemes across nine dimensions. The CLIP-ViL multimodal large model is a large AI model capable of simultaneously processing, understanding, and integrating multiple pieces of “modal” information, such as text, images, audio, video, 3D point clouds, and achieving cross-modal reasoning, generation, and interaction. The model’s pre-training data includes over 120,000 architectural design images (including subcategories of office, cultural, and industrial buildings), and its accuracy in extracting architectural professional features such as “window-to-wall ratio” and “mass dispersion” reaches 89.7%, significantly better than other models (≤72%). It supports customizing nine types of architectural scheme features, such as “color saturation” and “contour regularity”, through prompt engineering, which can be directly matched to the evaluation system of this study. Meanwhile, its open-source nature supports local deployment, enabling efficient batch evaluation of 90 schemes from three cases, avoiding the call restrictions and costs associated with closed-source models [17].

To ensure the objectivity of the evaluation results, the research design sequence adopted four steps [1]: data preprocessing, feature extraction, score quantification, and manual verification. First, unify the format of the generated scheme images, remove all other content such as text labels, etc., to avoid model bias, and rely on the model’s image information extraction capability to output raw scores. Secondly, use Z-score standardization to eliminate scale differences, and identify and correct three sets of outliers through box plots. Finally, verification by professional designers showed a correlation coefficient between manual and model scores of 0.87 (p < 0.001), indicating a low rate of misjudgment.

After the model identifies and quantifies the visual features of the images, its evaluation results are used for subsequent analyses. Subsequently, based on the sklearn toolkit, K-means cluster analysis is performed on the 30 retained schemes from the three cases. The optimal number of clusters is determined by the elbow method (where the sum of squared errors within clusters tends to stabilize), and the “Scheme Category Coverage” for each case is calculated (number of valid categories in the clustering result/total number of clusters) and compared with the diversity of traditional design (which generates an average of three schemes) [32]. Based on the K-means clustering algorithm in the sklearn toolkit, cluster analysis is performed separately on the 30 retained schemes from the three cases. After standardizing the original data as preprocessing, the optimal number of clusters is determined by the “elbow method”, i.e., performing clustering from two to ten classes for the data of the three projects (high-rise office, library, industrial warehouse building) and calculating the within-cluster sum of squared errors, selecting the inflection point in the line graph to determine the best number of clusters. To intuitively reflect the rationality of the clustering, the t-SNE (t-distributed Stochastic Neighbor Embedding) nonlinear dimensionality reduction technique is used to preserve the local similarity between data points, mapping the data from high-dimensional space to two-dimensional space. At the same time, the Principal Component Analysis (PCA) method is also used to map the high-dimensional data to two dimensions while retaining the maximum variance, resulting in the following two-dimensional scatter plots, respectively [33]. To further analyze the relationship between features and cluster categories, the clustered categories are used as pseudo-labels, and the Mutual Information (MI) method is used to calculate the features that contribute the most to distinguishing the category labels, excluding redundant features with high correlation, thus dividing the generated architectural design drawings into several major different styles [34].

2.4.4. Multi-Dimensional Evaluation

At the subjective level, a questionnaire was designed, targeting multiple dimensions such as “Scheme Innovation” and “Morphological Rationality”, inviting senior architects, owner representatives, and surrounding residents (non-architectural professionals) to score. At the objective level, building performance was simulated using relevant analysis software, such as EnergyPlus software to simulate the energy consumption and carbon emission characteristics of the schemes and Ecotect software to analyze the natural daylighting factor, quantifying sustainable performance indicators [35].

(1)

Regarding Questionnaire Design and Implementation

(a)

Questionnaire Conception: The content primarily focuses on the quality evaluation of design outcomes, aiming to compare the quality of renderings produced by AI-assisted design vs. the traditional workflow. According to the different focuses of the mid- and late stages, the mid-stage focuses on design expression and creative inspiration, while the late stage focuses on visual realism and Detail Completion. Referencing multiple relevant pieces of literature [1,36,37], such as the four core dimensions proposed by Albaghajati ZM et al. in “Exploring text-to-image application in architectural design: Insights and implications”, potential dimensional indicators were initially designed. Three frontline architectural designers (working in office, cultural, and industrial design, respectively, with over 15 years of experience) and two architectural theory research scholars (university professors) were invited for consultation and discussion to optimize the survey questionnaire, for example, by deleting non-core dimensional content such as construction feasibility, forming an evaluation system adapted to this case project.

(b)

Definition of Evaluation Dimensions and Corresponding Indicators: The specific content of the survey questionnaire is shown in Appendix A. The scoring mechanism follows a 7-point Likert scale, making the obtained data closer to continuous variables and the scale more scientific [37]. The specific scoring criteria are defined as 1 point = “Very poor/Completely disagree”, 2 points = “Poor/Basically disagree”, 3 points = “Slightly poor/Partially disagree”, 4 points = “Neutral/Neutral”, 5 points = “Slightly good/Partially agree”, 6 points = “Good/Basically agree”, and 7 points = “Very good/Completely agree” (for example, in the “Visual Realism” dimension, 1 point means “The image differs greatly from the real scene, lighting, shadows, and proportions are completely distorted”, and 7 points means “The image is close to a real photo, with no obvious deviations in lighting direction, material reflection, or perspective relationship”).

(c)

Reasons for choosing a 7-point scale instead of 5-point or 10-point scales are as follows: Data continuity—the normality distribution fit of data from a 7-point scale is significantly higher than that of a 5-point scale, allowing direct use in parametric statistical methods like independent samples t-test and ANOVA, meeting the need of this study to “quantitatively compare the differences between AI and traditional workflows”; secondly, regarding evaluation discrimination—5-point scales are prone to “central tendency stacking” (score 4 accounting for over 40%), while the 7-point scale refines the scoring gradient (e.g., the “Detail Completion” dimension can distinguish between “6 points (details are relatively rich, only a few components missing)” and “7 points (details are complete, no obvious defects)”), better fitting the professional need for “fine-grained evaluation” of renderings; additionally, the operational cost is controllable, as this scale can keep the completion time for a single questionnaire within 10–15 min, with a final effective recovery rate of 79.1% (meeting the academic survey standard of ≥70%).

(d)

Respondent Characteristics and Quantity Statistics: The survey was conducted on the final renderings for both the mid- and late stages, using online and offline distribution of subjective questionnaires. A total of 212 valid questionnaires were collected. The invited personnel were mainly senior architects from the company, owner representatives, and residents around the company. Specific respondent characteristics are as follows: 82 architectural designers (38.7%), among whom 45 were senior architects with ≥8 years of experience (accounting for 54.9% of the architect group, ensuring the authority of professional scoring); 45 project investors/owner representatives (21.2%, focusing on the practicality of outcomes); 57 non-architectural professional surrounding residents (26.9%, providing a public perspective); additionally, 28 academic researchers in the architectural field (13.2%, supplementing theoretical evaluation dimensions).

(e)

Supplementary Details on Survey Implementation: The survey period was from 20 February 2025 to 20 March 2025. At this stage, the late-stage scheme design of the cases was completed, ensuring that respondents could score based on the project’s final outcome drawings (non-simulated diagram). The distribution form adopted a mixed “online + offline” mode. Online distribution was through widely used local social and office software such as WeChat 4.1.5, QQ v9.9.23, and DingTalk 7.6.11. Offline distribution was conducted in research meeting rooms in Jinhua and Yiwu, owner’s office spaces, and near the author’s residence, distributing paper questionnaires (accompanied by color-printed renderings). A total of 268 questionnaires were distributed, and 212 valid questionnaires were recovered.

To verify the reliability of the data, this study used the SPSS AU V24.0 online statistical analysis platform. Based on the calculation formula, the scores for each dimension, each respondent’s scores for the mid- and late stages of the project, and the total scores were calculated according to the collected questionnaire data. Then, this statistical software was used to derive the Cronbach’s Alpha coefficient to verify the reliability level of the entire questionnaire [36].

(2)

Integration with Sustainable Development

The “sustainable development” mentioned in this study is not a broad environmental protection concept but rather performed to excavate the low-carbon potential throughout the life cycle at the architectural schematic design stage, with the linkage of “Form Parameters–Performance Indicators–Low-Carbon Goals” as the core, integrating sustainability requirements into the entire design process through quantitative tools, reducing environmental impact, and improving resource utilization efficiency. By extracting architectural form parameters such as floor area ratio, window-to-wall ratio, building shape coefficient, etc., and using professional simulation tools to evaluate the energy consumption, daylighting, and other performance of the schemes, design outcomes that combine creativity and low-carbon attributes are selected. Specifically, the covered dimensions are as follows:

(a)

Efficient use of natural resources. For example, using Ecotect software to evaluate the daylight factor of main functional areas (≥3%, complying with the “Standard for daylighting design of buildings” GB 50033-2013 [38]), calculating natural ventilation efficiency (≥0.7, complying with the “Assessment standard for green building” GB/T 50378-2019 [39]).

(b)

Controlled energy consumption and carbon emissions. For example, based on the building shape coefficient (≤0.40, complying with the “Design standard for energy efficiency of public buildings” GB/T 50189-2015 [40]), window-to-wall ratio (≤0.7, complying with the “General code for building energy efficiency and renewable energy utilization” GB/T 55015-2021 [41]), using EnergyPlus software to simulate the annual comprehensive energy consumption (Class A office building ≤65 kWh/m²·year, cultural building ≤55 kWh/m²·year, industrial building ≤50 kWh/m²·year, complying with the “Zhejiang Province Green Building Design Standard” DB33/1092-2021).

(c)

Adaptation of low-carbon materials and technologies. For example, focusing on the selection of building facade materials, collecting the proportion of low-carbon material usage in the design scheme (≥30% (based on the “Zhejiang Province Green Building Design Standard” DB33/1092-2021)); by determining the layout area of PV modules, solar collectors, etc., in the scheme design, using relevant green building assessment software (PKPM-GBS Green Building Design Software V3.3) to quickly calculate the contribution ratio of renewable energy.

In the traditional process, sustainability performance simulation often lags behind scheme finalization, whereas in generative AI-assisted design, sustainability performance analysis is front-loaded. The AI-generated concept sketches are used to extract key form parameters, such as floor area ratio, window-to-wall ratio, building shape coefficient, etc. The aforementioned relevant software is used for preliminary natural daylighting, ventilation, and energy consumption simulation to filter scheme directions with sustainable development potential [42]. Feasibility AI software, such as the Xiaoku AI platform, is also used to quickly quantify the morphological characteristics of the schemes and rapidly generate corresponding indicator data, avoiding disruptive modifications in later stages due to failure to meet indicators.

3. Results

3.1. Online Opinion Analysis Results

Based on the data from 80 high-attention articles on the topic of “AI and Architectural Design” crawled from Zhihu using Python3.10, this study obtained the general views of current Chinese architectural practitioners on the application effectiveness of generative AI. Except for the dimension of Technical Limitations and Practical Challenges, the average scores of the other seven dimensions were all above 6, indicating that the industry as a whole holds a positive and affirmative attitude towards the application of generative AI in architectural design. The average score for Design Efficiency and Process Optimization reached 8.28, certifying that most articles recognize the positive role of AI in the architectural field, such as AI simplifying the design process, improving design efficiency, and quickly generating alternative schemes. However, the average score for Technical Limitations and Practical Challenges was 5, showing a clearly neutral attitude among commentators. A considerable number of articles pointed out the limitations of AI in the architectural design field, such as data fragmentation, technical limitations, and talent scarcity.

This result clearly reveals a “dual perception” within the industry: on one hand, there is widespread optimism about the efficiency revolution and creative potential brought by generative AI; on the other hand, there is a wait-and-see attitude regarding the practicality and maturity of the technology for AI-assisted design projects. This cognitive perspective provides an important social background for the subsequent case studies in this research, highlighting the necessity of conducting effectiveness evaluation driven by data.

3.2. Application and Demonstration of Generative AI in the Early Stage of Design Projects

3.2.1. Demonstration Method: Efficiency Data Comparison

Compare the “Concept Generation Time” between traditional design (hand-drawing + SketchUp pro 2025 rough models) and Stable Diffusion-assisted design. Use descriptive statistics to present characteristic analysis and judge practical value, thereby verifying the time consumption comparison between the two control groups. The comparative data for the two control groups are shown in

Table 10. Comparison of concept generation efficiency in the early stage of architectural projects (unit: h).

Project Case	Traditional Design Time (Control Group A)	Stable Diffusion-Assisted Time (Experimental Group B)	Efficiency Improvement
	Mean ± SD (Median)	Mean ± SD (Median)	(Control Group A—Experimental Group B)/Control Group A × 100
Jinhua Information Economy Industrial Park	42 ± 4.5 (42)	2.8 ± 0.6 (3.0)	93.33%
Jinhua Library	45 ± 5.1 (45)	3.1 ± 0.7 (2.5)	93.11%
Yiwu Comprehensive Bonded Industrial Park	38 ± 3.2 (38)	2.3 ± 0.5 (2.4)	93.95%

.

In the early-stage concept generation phase of a project, the traditional process highly relies on manual experience and hand-drawn sketches, often constituting an efficiency bottleneck in the project cycle. The comparative test shown in Table 10 quantifies the impact brought by Stable Diffusion. Compared to the traditional hand-drawing and SketchUp pro 2025 rough modeling process, AI-assisted design drastically reduces the average design time from 38–45 h to 2.3–3.1 h, shortening the time required for architectural scheme generation by over 93%.

This result not only confirms the high expectations for “efficiency improvement” found in online opinions but also reveals the core advantage of generative AI from a data perspective: using diffusion models to efficiently compute abstract and vague requirements like text prompts, transforming them into a large number of concrete visual schemes, thereby greatly shortening the conversion path and time cost from “idea to visual scheme”.

3.2.2. Demonstration Method: Diversification Quantitative Analysis

The core of this section focuses on the practicality of generative AI in projects. To avoid redundant presentation of non-key content and ensure the focus of the discussion, this study only uses the traditional design schemes from Case 1 as typical representatives (

Table 11. Architectural design schemes by traditional design.

Category	Scheme Concept
Scheme 1 Conference Center and Office Building Combined
Scheme 2 Independent Conference Center along the Street
Scheme 3 Continuous Financial Institutions along the Street

), displaying the design effects output by the traditional group and the AI group, respectively. For the other two cases, only the results of the AI group are displayed, and the effects of traditional design schemes are not separately presented.

(1)

Jinhua Information Economy Industrial Park (Figure 7)

(2)

Jinhua Library (Figure 8)

(3)

Yiwu Comprehensive Bonded Industrial Park (Figure 9)

If the improvement in efficiency comes at the cost of sacrificing scheme diversity and innovation, its value would be greatly diminished. Therefore, this study utilizes the CLIP-ViL multimodal model to score the AI-generated schemes across nine dimensions—color saturation, contour regularity, detail richness, spatial hierarchy, mass dispersion, greening degree, light and shadow contrast, material complexity, and building void–solid relationship—comprehensively characterizing the architectural features of different schemes [1]. Below is a partial data sample (

Table 12. Sample of visualized dataset.

Case Category	Color Saturation (0–100)	Contour Regularity (0–100)	Detail Richness (0–100)	Spatial Hierarchy (0–10)	Block Dispersion (1–100)	Greening Level (0–100)	Light–Shadow Contrast (0–100)	Material Complexity (0–100)	Void–Solid Relationship (0–1)
High-Rise Office Building	65	85	78	7	15	30	80	85	0.82
High-Rise Office Building	68	88	82	7	25	28	82	88	0.78
High-Rise Office Building	60	82	75	6	10	32	78	80	0.86

), and style categories are objectively identified based on K-means cluster analysis.

As shown in the figure, the inflection point in each line chart indicates the appropriate k value. The k values for the three projects are 5, 4, and 5, respectively [43] (Figure 10).

In the three cases, points of the same color (samples of the same category) under both dimensionality reduction methods showed high spatial aggregation, meaning that points of the same category were closer in the reduced two-dimensional space, indicating that the clustering algorithm significantly separated the sample points (Figure 11).

Finally, for Project 1 (high-rise office building) and Project 2 (library), contour regularity and building void–solid relationship were selected for analysis. For Project 3 (industrial warehouse building), detail richness and building void–solid relationship were selected for analysis. The relevant scatter plots are as follows (Figure 12).

Regarding the sample data, for contour regularity, a higher number indicates a more irregular form; for the building void–solid relationship, a higher number indicates a higher proportion of void (transparency); for detail richness, a higher number indicates greater richness.

In the Jinhua Information Industrial Park project, contour regularity was divided into high, medium, and low parts with boundaries at 84 and 88; the building void–solid relationship was divided into low and high parts with a boundary at 0.775, thus classifying five styles. Red points exhibit high contour regularity and are primarily characterized by void walls, reflected in the architectural design drawings as simple, regular building forms with transparent and lightweight morphological features; pink points exhibit medium contour regularity and are primarily solid walls, with the design style characterized by interlocking blocks and distinct window–wall systems; gray points have rich forms with a large proportion of glass curtain walls, giving the building facade a sense of hierarchy; green points represent a free, dynamic, and sculptural modernist style; blue points exhibit the lowest contour regularity and the highest facade void-to-solid ratio, with a style possessing a sense of futurism and fluidity.

In the Jinhua Library project, contour regularity was divided into high, medium, and low parts with boundaries at 62 and 68; the building void–solid relationship was divided into low and high parts with a boundary at 0.55, classifying the generated architectural design drawings into five styles. Red points have low contour regularity and a high void–solid ratio, presenting features of square, regular forms and transparent, open facades; pink points have medium contour regularity and a low void–solid ratio, with a style characterized by interlocking blocks and dominant solid walls; green points have medium contour regularity and a high void–solid ratio, with building forms that are rich and varied, and facades that are open and lightweight; blue points have high contour regularity and a low void–solid ratio, forming complex geometric shapes with a strong sculptural sense.

In the Yiwu Comprehensive Bonded Industrial Park project, detail richness was divided into low, medium, and high parts with boundaries at 26 and 31; the facade void–solid ratio was divided into low and high parts with a boundary at 0.3, generating architectural concept drawings of five styles. The style tendency of blue points is extremely simple forms and relatively enclosed facades, representing ultimate purity; the architectural style classified into green points involves regular rectangular volumes with continuous large high windows or strip windows; pink points have simple forms but facades rich in material variation; gray points have facades rich in detail, transparent and open, with a style leaning towards high-tech and futurism; red points have complex forms and solid, heavy facades, highlighting the structural strength and mechanical esthetics of industrial architecture.

Through data analysis, it can be seen that in the three cases, the “category coverage” of AI-generated schemes is five style types for office building design, four style types for library design, and five style types for industrial park design. This diversity is significantly higher than the number of styles per scheme set in the traditional design process. Secondly, through t-SNE and PCA dimensionality reduction visualization, schemes of different styles show clear clustering distribution in the two-dimensional space, indicating that AI can systematically expand design styles rather than producing random, disordered variations.

In summary, generative AI not only achieves an order-of-magnitude improvement in generation efficiency but also, through its powerful feature learning and combination capabilities, broadens the combinatorial boundaries of design schemes. From the cluster analysis results, the significant style categories possessed by AI-generated schemes can provide design teams with systematic creative options far beyond traditional methods, thereby enabling more comprehensive coverage of potential design directions in the early stages of a project.

3.2.3. Preliminary Assessment of Sustainability Indicators

Using Project 1 (Jinhua Information Economic Industrial Park) as an example for demonstration, after the AI-generated concept schemes, their key form parameters were extracted, and rapid energy consumption and daylighting analysis were conducted within performance simulation software for a preliminary assessment of sustainability indicators. The data indicators are as follows (

Table 13. Sustainability performance analysis of partial schemes in Project 1.

Code	Window-to-Wall Ratio	Building Shape Coefficient	Facade Design Strategy	Average Daylight Factor in Main Functional Areas (%)	Estimated Annual Comprehensive Energy Consumption (KWh/m2)
A-1	0.58	0.30	performance insulating LOW-E glass curtain wall	3.7	66.3
A-2	0.50	0.33	Comprehensive external shading + metal curtain wall	3.1	62.1
A-3	0.62	0.29	Double-skin facade (breathing facade)	4.0	64.8
A-4	0.47	0.35	Vertical shading grille + point windows	2.8	59.5
A-5	0.53	0.31	Glass curtain wall + integrated vertical PV panels	3.5	63.4
Reference Standard	≤0.70	≤0.40	-	≥3.0	≤65

).

According to the data in the table, the estimated sustainability performance indicators for schemes A-2, A-3, and A-5 meet the requirements, while schemes A-1 and A-4 are less reasonable and can be excluded or optimized. For the initial schemes of a project, this characteristic of “rapidity” and “relevance” is a great advantage of generative AI. It can provide a data basis for the rapid screening and preliminary evaluation of sustainability indicators in the early stages, while also helping the design team effectively identify the relationship between “form parameters” and “sustainable performance”, avoiding adjustments later due to failure to meet indicators and filling the gap between “scheme conceptualization” and “indicator quantification” in traditional design. In the initial scheme presentation of this project, the AI-assisted design group included sustainability indicators in the presentation material, which was recognized by the owner.

3.3. Mid- and Late-Stage Application Effectiveness Results

3.3.1. Demonstration Method: Efficiency Analysis

The decision-making efficiency of the two groups was compared, recording the “Scheme Deepening Cycle” and the “Number of Meeting Rounds Required to Reach a Decision” for both the traditional design group and the Stable Diffusion-assisted design group. The comparison data for the three cases are shown in

Table 14. Comparison of project deepening cycle and decision-making rounds.

Project Case	Single Traditional Design Deepening Cycle (Hours)	Stable Diffusion Assisted Deepening Cycle (Hours)	Traditional Design Decision Rounds	Stable Diffusion-Assisted Decision Rounds
Jinhua Information Economy Industrial Park	40–56	12	4	2
Jinhua Library	24–40	9.6	5	2
Yiwu Comprehensive Bonded Industrial Park	24–40	14.4	4	2

and Figure 13.

For the mid-stage of the project, the traditional process relies on re-modeling in SketchUp pro 2025 and rendering in Lumion, with a single deepening iteration taking an average of 5 days. Comparing multiple schemes requires additional modeling, leading to 4–5 decision rounds. In contrast, the core requirement of the AI-assisted design process is “rapid response to modifications + multi-scheme comparison”. In these three projects, the group using Stable Diffusion for assisted design shortened the scheme deepening cycle by 78.33%. This is because a single deepening iteration produces a larger number of schemes, providing a wealth of visually comparable options, which reduces communication costs and accelerates the decision-making process. Furthermore, generative AI, through the “image-to-image” function combined with control plugins like ControlNet and LoRA, enables refined rapid iteration and precise variants of existing schemes, thereby significantly reducing the number of decision rounds and improving mid-stage decision efficiency.

In the late stage of the project, the traditional group handed over the refined model of the architectural scheme to a post-production rendering company for effect optimization. The AI generation group used SketchUp pro 2025’s built-in Inspiration Rendering plugin for real-time precise rendering of the scheme. After multiple adjustments, the final scheme results were obtained (Figure 14), and their production cycle and cost were recorded (

Table 15. Comparison of drawing time and cost.

Evaluation Dimension	Traditional Design (Rendering Company)	AI-Assisted Design with Inspiration Rendering Plugin
Single-Image Production Cycle (Hours)	48 ± 6	1 ± 0.5
Single-Image Cost (CNY)	1200 ± 200	100 ± 10 (Labor and Hardware Energy Cost)

).

According to the records, compared to the traditional rendering company production model, AI-assisted design achieved extremely high efficiency and cost optimization in the late-stage result output phase. The design team could perform real-time scheme modifications and re-presentation based on feedback during presentations or decision-making meetings. In terms of cost, the cost per image dropped from a four-digit figure to a two-digit figure, completely changing the traditional notion of renderings as “high-value, low-frequency deliverables”, transforming them into a high-frequency, low-cost iterative design tool. This order-of-magnitude change enables small design firms or design teams with tight project budgets to obtain rich visual results at a very low cost, greatly enhancing their competitiveness in the market.

3.3.2. Demonstration Method: Output Quality Analysis

In judging application effectiveness, the evaluation of the output quality dimension is also a crucial part. The questionnaire survey results provide an important balancing perspective to this extremely high-efficiency data. The overall Cronbach’s α coefficient of the questionnaire was 0.876, and in the dimension of Detail Completion, Cronbach’s α coefficient was 0.912 (

Table 16. Questionnaire reliability.

	Analysis Scope	Number of Items	Sample Size	Cronbach’s α
Mid-stage	Clarity of Design Concept Expression	3	212	0.792
	Spatial and Volumetric Relationship	3	212	0.886
	Overall Atmosphere and Esthetic Value	3	212	0.908
	Potential to Inspire Deepening Design	2	212	0.898
Late stage	Visual Realism	3	212	0.905
	Detail Depiction and Completion	3	212	0.912
	Material and Texture Representation	2	212	0.904
	Image Artistry and Harmony	3	212	0.807

), indicating high consistency in the questionnaire’s comparison of preferences regarding image depth and quality, and the evaluation results are reliable.

According to the analysis of the survey results (Figure 15 and Figure 16), the public has a relatively high understanding of AI, with 70% of architectural practitioners having used AI-assisted design in some projects. In the mid-stage of the project, the AI-assisted generated renderings performed best in the dimension of “Potential to Inspire Deepening Design”, especially the “Creative Inspirativeness” indicator (AI mean: 5.6 ± 0.5 vs. Traditional mean: 4.0 ± 0.7, t = 18.24, p < 0.001). This strongly aligns with the core need of “scheme divergence–refinement” in the mid-stage. Therefore, it is recommended that more architectural practitioners use generative AI to assist team design during the mid-stage scheme exploration phase of design projects. In the late stage of the project, the traditional workflow scored more prominently. The traditional workflow received significantly higher scores than AI in dimensions such as “Visual Realism”, “Detail Completion”, and “Material and Texture Representation” (p < 0.05). This clearly delineates the current capability boundaries of generative AI technology in late-stage applications. It is an excellent “rapid scheme explorer” and “communication assistant”, but there is still a certain gap between achieving final deliverable-level results through this technology—such as photorealistic quality, accurate detailed construction expression, and complex physical light and shadow simulation—and the capabilities of experienced rendering engineers and mature manual rendering processes. At this stage, generative AI cannot yet fully replace them. In the future, generative AI still has great potential for technological improvements in the construction industry. Issues such as potential logical errors in details, material distortion, and perspective deviation are key technical bottlenecks that need continuous optimization.

4. Research Discussion

4.1. Generative AI Technology Reshaping Architectural Workflows

Based on empirical data from design projects, the reshaping of architectural design workflows by generative AI is reflected in three core dimensions: “Process Paradigm Transformation”, “Data-Driven Empowerment”, and “Front-Loading Sustainable Design”, forming a new design model of “Human–Machine Collaboration”.

Regarding the Process Paradigm Transformation dimension, it shifts from the traditional “Linear Progression” to the “Cyclical Iteration” of generative AI. Traditional architectural design processes follow a linear logic of early, middle, and late stages, progressing step by step with high rework costs. The use of generative AI can break this linear logic, adopting a cyclical paradigm of “Rapid Generation–Quantitative Evaluation–Instant Iteration” [44]. In the early stage, multiple schemes are generated using AI model software like Stable Diffusion, and usable schemes are selected based on cluster analysis and sustainability evaluation results. In the middle stage, based on client requirements, new design drawings are quickly generated using AI for scheme modifications, avoiding the long rework cycles associated with purely manual design in traditional processes. This paradigm transformation increases the “response speed” of architectural design by 3–5 times, significantly enhancing the flexibility of the design process.

Regarding the Data-Driven Empowerment dimension, the application of generative AI in architectural design promotes a shift from experience-dependent dominance to quantitative decision-making. In the early conceptual scheme stage, cluster analysis provides quantitative indicators for design scheme categories, avoiding the relatively fixed and singular design approaches in traditional processes. In the mid-stage deepening phase, the use of T-test results and reliability/validity analysis provides statistical support for decisions made by clients and design teams. In the later stage rendering phase, quantitative scoring from double-blind tests reveals the shortcomings of AI in detailed rendering performance and directions for improvement.

Regarding the Sustainable Design Front-Loading dimension, in the traditional model, “post-optimization” of energy efficiency after the scheme is finalized leads to limited optimization space. In contrast, generative AI, through the early intervention of energy consumption simulation software, directly guides the scheme towards low-carbon, avoiding rework later due to failing energy consumption data, thus promoting the low-carbon transformation of the construction industry.

4.2. Limitations and Future Work

4.2.1. Limitations

Regarding technical limitations, the following issues exist:

(1)

Limitations of the questionnaire. Considering operability, the questionnaire content for mid-to-late-stage architectural design should not be excessive. The comparative analysis was only representative, using “high-rise office” type projects, resulting in slightly less coverage of project types. Future work needs to further verify performance for other project types.

(2)

Hardware acceleration limitations. Commonly used large generative AI models like Stable Diffusion require high-performance GPUs to generate high-resolution (2048 × 2048) images, resulting in high hardware costs. Furthermore, the refinement and realism need improvement. Currently, this poses an operational barrier for small architectural firms lacking professional computing infrastructure.

(3)

Insufficient interoperability within the software ecosystem. Currently, apart from Inspiration Rendering being an internal plugin for SketchUp pro 2025, allowing interoperability and real-time rendering, the integration of mainstream generative AI models with industry-standard BIM and CAD platforms (e.g., Revit and ArchiCAD) is limited. This hinders the seamless flow and real-time iteration of design data, constraining the scalability of technology application.

(4)

Insufficient Domain-Specific Knowledge Embedding: Current generative models lack embedded architectural code compliance checks and structural logic, requiring manual post-processing and external validation for regulatory adherence and constructability.

(5)

Prompt Sensitivity and Operational Threshold: Output quality and semantic alignment are highly sensitive to prompt engineering and parameter tuning (CFG Scale, Denoising Strength, ControlNet module selection, etc.), necessitating advanced user expertise and increasing operational complexity.

(6)

Data Provenance and Copyright Risk: Training datasets sourced from publicly available architectural imagery may introduce copyright infringement risks and high-similarity outputs, necessitating improved dataset curation and provenance tracking [16].

4.2.2. Future Work

For the entire technical field, future AI technology should develop towards three directions: “Deep Integration”, “Customized Modules”, and “Full Process Coverage”. Develop multi-dimensional collaboration with software like BIM (Building Information Modeling) and AR (Augmented Reality)/VR (Virtual Reality) to achieve interactive design modes where AI schemes can be updated, modified, and experienced immersively in real-time; achieve high integration with architectural knowledge to improve the accuracy of sustainable design; ensure AI coverage of the entire architectural process and add related content such as quality assessment to ensure the rational development of AI architectural generation technology [45]. For practitioners in the architectural industry, they should keep up with research progress in AI-related fields, combine this technology with architectural design as maximally and scientifically as possible, and promote the deep application of generative AI in this field.

5. Conclusions

This study, centered on generative AI technologies such as Stable Diffusion and Inspiration Rendering, and based on the application of existing large generative AI models and industry development trends, combined with the author’s firsthand experience in three design projects within Chinese architectural design institutes, conducted research on the application effectiveness of generative AI from data-driven and sustainable perspectives. It proves that generative AI has good applicability in the early and middle stages of design, significantly optimizing the entire process efficiency of projects. Specifically, early-stage concept generation time was shortened by over 90%, scheme diversity coverage increased by over 55%, and sustainable design achieved “source injection”.

Although current generative AI technology has many areas for improvement, practice has proven that these issues can be gradually resolved through means such as the development of customized models and model iteration and upgrades. For small and medium-sized enterprises, they can leverage AI to assist in early-stage concept generation and design deepening, quickly obtaining “technological dividends”. For large enterprises, they can gradually build internal, independent “AI Scheme Databases” to form differentiated competition. For university education, it is recommended to add relevant courses on “AI-assisted Applications” and strengthen the integrated teaching of “AI + Sustainable Design”, guiding students to use tools effectively and focus more energy on the essential issues of architectural design [46].