T2I generation in architecture

T2I AI tools – driven by diffusion models – enable designers to create visual architectural concepts based on natural language prompts. These models, trained on large datasets of images and captions, have been widely adopted for conceptualization, ideation, and rapid prototyping in architectural design (Hanafy, Reference Hanafy2023; Horvath and Pouliou, Reference Horvath and Pouliou2024; Paananen et al., Reference Paananen, Oppenlaender and Visuri2024). Research has demonstrated their usefulness in visual storytelling, early-stage façade studies, and atmospheric renderings (Çelik, Reference Çelik2024; Montenegro, Reference Montenegro2024). However, critical investigations into their spatial logic, functional accuracy, or environmental responsiveness are limited.

Despite their potential, current T2I systems often prioritize aesthetic coherence over structural realism or performative intelligence (Iranmanesh and Lotfabadi, Reference Iranmanesh and Lotfabadi2024). Some recent studies (e.g., Liao et al., Reference Liao, Huang, Zheng and Lu2022) suggest combining these outputs with CAD and BIM (Building Information Modeling) tools to evaluate architectural feasibility. This hybrid workflow – prompt to image to simulation – is emerging as a method to assess the generative output’s alignment with real-world architectural standards.

Recent work has explored text-to-image generative models (e.g., Midjourney, DALL·E, and Stable Diffusion) as design tools for early-stage architectural ideation. These models allow architects to rapidly visualize multiple concept sketches from written prompts, effectively broadening the range of design options. For example, Paananen et al. (Reference Paananen, Oppenlaender and Visuri2024) found that using Midjourney, DALL·E, and Stable Diffusion in a classroom exercise supported “serendipitous discovery of ideas and an imaginative mindset,” enriching the concept phase of a cultural center design. Similarly, a recent study reports that T2I tools help architects “conceptualize new architectural ideas more clearly,” providing fresh perspectives and expanding creativity by tapping into large visual datasets (Thampanichwat et al., Reference Thampanichwat, Wongvorachan, Sirisakdi, Chunhajinda, Bunyarittikit and Wongmahasiri2025). In practice, leading design firms already use these AIs for divergent-thinking exploration: Researchers note that models like Midjourney, DALL·E, and Stable Diffusion “promote rapid exploration and iteration through visualization, enabling designers to better express their design concepts” (Chen et al., Reference Chen, Song, Guo, Sun, Childs and Yin2025). These findings suggest that text-based image generators can become meaningful parts of the design workflow, especially when prompts and constraints are carefully managed (although attention must be paid to their limitations).

Daylight simulation in architectural design

Daylight simulation tools, including VDV, Radiance, and ClimateStudio, have long been utilized to predict and optimize natural light penetration within built environments (Kim and Chung, Reference Kim and Chung2011; Sancho-Salas et al., Reference Sancho-Salas, Flor and Muñoz2023). These tools evaluate daylight autonomy, uniformity, and glare to guide decisions on window orientation, room depth, material reflectance, and spatial configuration (Cammarano et al., Reference Cammarano, Pellegrino, Lo Verso and Aghemo2015; Gibson and Krarti, Reference Gibson and Krarti2015). Among them, VDV provides a user-friendly interface and accurate simulation outputs based on geographic data and seasonal solar positions.

Computational modeling and simulation play a critical role in evaluating the performance of architectural designs. When these tools are embedded within the design workflow, the process is referred to as performance-based design. Although rapid feedback loops are essential throughout design development, conducting daylight simulations remains a complex task due to their intensive computational demands and time-consuming nature (Queiroz et al., Reference Queiroz, Fernandes and Pereira2024). Performance-based design using such tools is standard practice in passive design strategies, especially in climates where lighting energy demand is critical (Oxman, Reference Oxman2007). However, their application to AI-generated content is novel. No known study has systematically simulated T2I-generated plans for daylighting performance across multiple climate zones – a gap this research seeks to address.

Daylight performance is now usually assessed with dynamic, climate-based simulation rather than static rules of thumb. In the past, simple metrics like the Daylight Factor under a single overcast sky were common; however, climate-based daylight modeling (CBDM) has gained prominence (Jiang, Reference Jiang2021). CBDM uses hourly weather data (latitude, climate, orientation, etc.) to predict interior luminance over the year, accounting for solar geometry and sky variability. This approach “has attracted considerable attention” and is gradually replacing fixed-condition methods (Jiang, Reference Jiang2021). Accordingly, researchers have adopted both advanced simulation engines (Radiance and Daysim) and industry tools like the VDV. VDV 3.0 (a free Radiance-based program) is widely used for building-scale analyses. For instance, Mandala et al. (Reference Mandala, Sutanto and Santoso2021) used VDV 3.0 to model a large-volume building with various skylight designs and computed daylight factors and uniformity across the space.

In practice, standards often call for checking daylight at equinox conditions to represent average seasonal performance. For example, LEED (Leadership in Energy and Environmental Design) lighting rules instruct teams to simulate illuminance at 9:00 and 15:00 h on March 21 (or September 21) to gauge daylight availability. In one study, a classroom testbed in India was modeled in VDV 3.0 and simulated on March 21 (equinox) from 8:00 to 18:00 h; comparing modeled results to on-site measurements showed that using an accurate sky model (based on prevailing weather data) improved prediction of work-plane illuminance by about 24% (Edwards and Torcellini, Reference Edwards and Torcellini2002; Boyce, Reference Boyce2004). Such studies illustrate that software like VDV can reliably compute time-resolved illuminance for design evaluation. Thus, the literature emphasizes climate-based, hourly simulation (e.g., Spatial Daylight Autonomy and Useful Daylight Illuminance) and the use of realistic sky models and metrics, rather than single-point calculations.

AI-driven climate-based housing design

Several recent projects have experimented with AI-generated architectural solutions tailored to climatic demands. Recent research has highlighted the potential of AI and data mining techniques in addressing the challenges of energy poverty (EP), particularly in warm climate zones. According to Bienvenido-Huertas et al. (Reference Bienvenido-Huertas, Sánchez-García, Marín-García and Rubio-Bellido2023), AI-based automated tools can effectively detect EP without the need for detailed energy performance analyses. This approach not only streamlines the workload of energy managers and social workers but also enhances the predictive capacity of EP diagnosis, paving the way for more efficient and scalable interventions.

In the context of smart environmental management, the integration of AI with Internet of Things technologies offers promising advancements in greenhouse monitoring systems. As highlighted by Riskiawan et al. (Reference Riskiawan, Gupta, Setyohadi, Anwar, Kurniasari and Hariono2024), traditional greenhouse environments still largely depend on manual regulation of temperature and humidity, which imposes labor-intensive demands. Their study demonstrates that AI-enabled automated environmental control significantly enhances the precision and efficiency of indoor climate regulation. The system’s capacity to autonomously predict and adjust environmental conditions marks a substantial shift toward intelligent and sustainable greenhouse practices.

The intersection of generative AI and environmental simulation presents new opportunities to evaluate AI’s design “intelligence.” This article builds upon emerging discourse by combining T2I models with simulation software to assess the daylight performance of climate-specific house plans. It also contributes to the methodological expansion of architectural research by incorporating computational workflows that link creativity and performance.

Researchers are beginning to combine AI generative design with climate data to create performance-aware housing layouts. Generative AI can produce novel floorplan alternatives, but climate adaptation requires coupling generation with environmental evaluation. Recent reviews highlight that many existing AI floorplan tools implicitly depend on local weather data and, therefore, yield solutions tailored to specific regions (Meselhy and Almalkawi, Reference Meselhy and Almalkawi2025). For true regional adaptation, climate must be explicitly integrated: Different climate zones impose different design priorities (e.g., maximizing shading and minimizing solar gain in hot-arid regions versus enhancing natural ventilation and passive heating in temperate zones) (Meselhy and Almalkawi, Reference Meselhy and Almalkawi2025). Meselhy and Almalkawi (Reference Meselhy and Almalkawi2025) note that without incorporating these factors, a design optimized for one climate “may not be directly applicable elsewhere” unless adapted to new climate inputs (Meselhy and Almalkawi, Reference Meselhy and Almalkawi2025).

In response, some recent studies use AI pipelines that generate floor plans and simultaneously predict performance metrics from climate data. For example, Hu et al. (Reference Hu, Zheng and Lai2024) developed a workflow where a diffusion model generates a variety of residential floor plan layouts, and a neural network (GAN, Generative Adversarial Network) rapidly predicts each layout’s daylight performance (Hu et al., Reference Hu, Zheng and Lai2024). They fine-tuned the model on regionally realistic house plans and then used the GAN as a surrogate to evaluate daylight autonomy. Their AI workflow reproduced daylight simulation results with high fidelity (within ~5% error of ground-truth Radiance simulations) and ran over 200× faster than traditional iteration (Hu et al., Reference Hu, Zheng and Lai2024). This demonstrates that AI-generated designs can be quantitatively assessed for daylight even at an early stage. Other efforts similarly integrate climate: for example, surrogate modeling studies show that including detailed hourly weather features (temperature, solar angles, etc.) improves prediction accuracy across diverse climate zones (Manmatharasan et al., Reference Manmatharasan, Bitsuamlak and Grolinger2025) (suggesting future AI tools should ingest location-specific climate datasets). In sum, the literature indicates that AI methods for housing design are starting to account for environmental performance: Generative models propose geometry, then machine-learning predictors (or embedded simulations) assess metrics like daylight, allowing architects to iteratively refine climate-adapted designs.

Research design

This study adopts a mixed-methods design combining generative AI modeling, digital drafting, and environmental simulation to evaluate the daylighting performance of AI-generated sustainable housing plans across diverse climate zones. The workflow is structured into three main phases: (1) Plan generation via T2I diffusion models, (2) plan reconstruction in AutoCAD for simulation compatibility, and (3) daylight analysis using VDV based on equinox conditions. To ensure methodological clarity, the workflow is further divided into seven stages, including climate zone and city selection, prompt engineering and AI tool evaluation, two- (2D) and three-dimensional (3D) modeling processes, and multi-date daylight simulations using false color mapping. Evaluation criteria are based on average illuminance values (lux) in key interior spaces, benchmarked against recognized standards from the literature. The comparative analysis highlights the extent to which generative models consider passive lighting design principles during the early stages of architectural production (Table 1).

Table 1. Workflow stages of the AI-based climate-sensitive housing study

Climate zone and city selection

To ensure climatic diversity, five cities were selected from distinct Köppen-Geiger climate zones (Peel et al., Reference Peel, Finlayson and McMahon2007):

• • Jakarta, Indonesia (Af) – Tropical rainforest climate

• • Alice Springs, Australia (BWh) – Hot desert climate

• • Madrid, Spain (Csa) – Warm-summer Mediterranean climate

• • Winnipeg, Canada (Dfb) – Cold continental climate

• • Tromsø, Norway (ET) – Polar tundra climate

The selection criteria included the following: (a) Representativeness of major global climatic typologies, (b) urban settings with housing demand, and (c) availability of geographic and solar data for simulation.

To ensure comparability across climate zones and isolate the influence of environmental context on model behavior, a standardized prompt was used for each location. This decision was informed by the need to minimize confounding variables arising from prompt variation, which can significantly affect generative outputs. The single-prompt approach enabled a more controlled assessment of model performance under varying climatic conditions.

Prompt design and AI model selection

Architectural floor plans were generated using leading open-source T2I diffusion models through high-resolution inference, incorporating architecture-specific fine-tuning. In this study, three distinct AI tools were selected to generate sustainable housing plans tailored to various climate zones: ChatGPT (OpenAI), Microsoft Copilot Image Creator, and LookX AI. Each tool offers unique advantages in architectural plan generation, and their selection was based on criteria such as accessibility, architectural specialization, and technical capability.

• • ChatGPT (OpenAI): Developed by OpenAI, ChatGPT integrates the DALL·E 3 image generation model within the GPT-4o multimodal framework. This enables T2I conversion based on user-defined prompts. In this study, ChatGPT served a dual role as both a prompt engineering assistant and a generative model, producing visual outputs derived from its own textual descriptions. The multimodal capacity of GPT-4o allows seamless interaction between textual and visual content, facilitating the generation of high-quality, context-sensitive architectural images.

• • Microsoft Copilot Image Creator: Offered through Microsoft’s Copilot platform, this tool utilizes Bing Image Creator technology to transform textual prompts into photorealistic images. Distinguished by its free and open-access availability, user-friendly interface, and rapid rendering capability, Copilot efficiently processes natural language inputs to produce high-quality visuals, making it an accessible solution for early-stage generative design tasks.

• • LookX AI: Specifically designed for architecture and interior design applications, LookX AI converts both text-based and sketch-based inputs into high-resolution renderings. The platform caters to design professionals by enabling custom model training and stylistic output generation. By streamlining the integration of AI into the architectural design workflow, LookX supports creative exploration and iterative visualization in professional contexts.

Each city was assigned a tailored prompt that included:

• • climate-adaptive design features (e.g., elevated structures, thermal mass, insulation, and glazing orientation),

• • material specifications (e.g., bamboo, adobe, timber, and concrete),

• • spatial configuration (one bedroom, one kitchen, and one living room),

• • bioclimatic and passive design principles (e.g., cross-ventilation, shading devices, and courtyard inclusion).

Prompt engineering focused on maximizing climatic specificity and functional clarity to encourage AI-generated layouts with architectural intent. Below are the final versions of the climate-based prompts used in this study:

• • Jakarta, Indonesia (Tropical climate):

  Tropical climate sustainable house plan in Jakarta, Indonesia. Elevated structure with natural ventilation. Large operable windows for cross-ventilation, positioned to maximize daylight. Open floor plan with semi-outdoor living spaces. Locally sourced materials such as bamboo, teak, and recycled wood. Biophilic design with lush greenery, internal courtyards, and shading devices. Floor plan must include one bedroom, one kitchen, one living room, and one toilet. The design must prioritize airflow, cooling efficiency, and sustainable urban living.

• • Alice Springs, Australia (Desert climate):

  Desert climate sustainable house plan in Alice Springs, Australia. Thick adobe or rammed earth walls for thermal mass. Small, strategically placed windows with shading devices to reduce heat gain. Courtyard layout to allow passive cooling and outdoor living. Minimal openings on west-facing walls. Roof insulation and solar panels integrated. The floor plan must include one bedroom, one kitchen, one living room, and one toilet. Locally sourced materials and efficient water usage features should be included.

• • Madrid, Spain (Mediterranean climate):

  Mediterranean climate sustainable house plan in Madrid, Spain. Thick masonry walls for thermal stability. Shaded verandas and balconies to manage solar gain. Cross-ventilation through aligned operable windows. Sloped tile roof for seasonal adaptability. Local stone and timber used. One-bedroom, one-kitchen, one-living room, one-toilet layout with semi-open spaces like patios or loggias.

• • Winnipeg, Canada (Cold continental climate):

  Cold climate sustainable house plan in Winnipeg, Canada. Compact form with high insulation levels. Triple-glazed south-facing windows to maximize solar heat gain. Minimal window openings on north walls. Airtight construction with passive solar principles. Locally sourced timber and insulated concrete. One-bedroom, one-kitchen, one-living room, and one-toilet layout. Enclosed vestibule to prevent heat loss.

• • Tromsø, Norway (Polar climate):

  Polar climate sustainable house plan in Tromsø, Norway. Elevated foundation to address snow buildup. Super-insulated envelope with minimal thermal bridging. South-facing windows with deep frames. Airtight design with mechanical ventilation and heat recovery. Compact one-bedroom, one-kitchen, one-living room, one-toilet layout. Use of sustainable local timber and snow-shedding roof form.

These prompts served as standardized instructions to assess whether AI-generated plans integrated region-specific climate strategies and spatial logic.

The floor plans generated by the three selected AI tools were developed based on previously defined climate-specific prompts. These outputs were subsequently transferred to the AutoCAD environment and subjected to daylight performance simulations using VDV. This workflow enabled a comparative assessment of the extent to which each AI model integrated climate-adaptive design strategies into spatial planning.

However, not all outputs produced by the AI tools were equally suitable for architectural drafting. The results generated by LookX AI were excluded from the CAD modeling phase due to the absence of windows or other fenestration elements, as well as poor legibility from a plan-reading perspective. Similarly, image outputs for two cities produced by Microsoft Copilot Image Creator were deemed unfit for AutoCAD translation, as they lacked architectural clarity and instead exhibited highly abstract forms that could not be interpreted as floor plans. Consequently, the CAD drafting and daylight simulation stages of the study were carried out exclusively using the plans generated by ChatGPT (OpenAI) and valid outputs from Microsoft Copilot. This selective process underscores the importance of evaluating the consistency and architectural usability of AI-generated floor plans, particularly in climate-responsive design contexts.

While this study focused on three AI models selected for their accessibility and architecture relevance (ChatGPT/DALL·E, Microsoft Copilot, and LookX AI), it must be acknowledged that the rapidly expanding ecosystem of generative tools, including Midjourney and Stable Diffusion, offers additional capabilities that may lead to different architectural outcomes. The limited model diversity is thus recognized as a methodological limitation.

Plan extraction and CAD translation

Since AI-generated images were raster-based and lacked precise metric information, all floor plans were manually redrawn and vectorized using AutoCAD 2025 to ensure simulation compatibility and dimensional consistency. The translation process followed these standard conventions:

• • Wall thicknesses were standardized across all plans as follows: 20 cm for external walls, and 15 cm or 10 cm for internal partitions, depending on the wall typology suggested in the AI output.

• • Window and door placements were preserved as visible in the AI-generated plans to maintain spatial intent and ventilation logic.

• • Each plan was scaled to 1:50, and north orientation was maintained.

This CAD translation process enabled consistent geometric input for subsequent daylight simulation and comparative analysis across different climate-adaptive designs.

Following the drafting phase, a 3D model of each plan was constructed in AutoCAD to prepare for daylight simulation. In this stage:

• • Architectural elements, such as floors, walls, ceilings, windows, and doors, were assigned to separate layers.

• • This stratification allowed for accurate material assignment in the VDV environment, as the simulation software interprets geometry and materials based on imported CAD layer configurations.

This modeling step provided the necessary semantic and geometric clarity for translating AI-generated architectural intent into scientifically valid simulation scenarios.

Daylight simulation with VDV

After 3D modeling in AutoCAD, each plan was exported to VDV (version 3.0) to assess natural light performance under varied solar conditions. VDV is a validated daylighting analysis tool based on the Radiance rendering engine. While a range of daylight simulation engines exists, such as Radiance, DIVA, and Honeybee, VDV was selected due to its intuitive user interface, high compatibility with AutoCAD models, and suitability for residential-scale analysis. Although not as customizable as Radiance-based engines, it enables efficient and replicable simulations for comparative purposes within the scope of this study.

Simulations were conducted using Parameters included:

• • Location settings: City-specific geographic coordinates (latitude and longitude)

• • Time of year: March 21 and September 21 (equinoxes), June 21 (summer solstice), and December 21 (winter solstice)

• • Sky model: CIE overcast and intermediate sky types

• • Simulation outputs: Illuminance levels (lux)

• • Material reflectance values: Assigned based on realistic assumptions about locally sourced materials inferred from the AI prompt. For example:

  • ○ Walls: Reflectance of 0.50–0.65 depending on surface finish (e.g., wood, adobe, and concrete)

  • ○ Floors: Reflectance of 0.30–0.45

  • ○ Windows: Transparent glazing with visible transmittance of 0.70

  • ○ Ceilings and shading elements: Matte materials with reflectance below 0.25, where applicable

The 3D models were prepared with layered materials in AutoCAD to facilitate proper material mapping within Velux. Each building element (walls, windows, and floors) was placed on a distinct CAD layer and assigned corresponding material properties upon import.

Analysis parameters included:

• • Rendered daylight scenes at 12:00 h (local solar time) for each simulation date,

• • Generation of false-color illuminance maps for internal spaces.

This simulation framework enabled the objective evaluation of how well AI-generated plans performed under natural daylight scenarios, revealing whether the T2I model implicitly considered daylight access during design generation.

Evaluation criteria

The daylight performance of each AI-generated housing plan was evaluated based on the average indoor illuminance (lux) values obtained from simulations conducted on equinoxes (March 21 and September 21) and solstices (June 21 and December 21). The analysis focused on three primary functional spaces: the living room, the kitchen, and the bedroom.

According to commonly accepted daylighting performance standards (Illuminating Engineering Society, 2011; Mardaljevic, Reference Mardaljevic2000), the following average illuminance thresholds were applied: 300–500 lux for living rooms, 500–750 lux for kitchens, and 200–300 lux for bedrooms. These benchmarks are widely recognized in lighting design literature and serve as a basis for evaluating the sufficiency of natural light in residential settings.

By comparing the simulated illuminance values with these standards, the study aimed to determine whether the AI-generated designs inherently considered adequate daylighting strategies suitable for their respective climatic contexts.

Architectural evaluation of AI-generated plans

In the initial phase of the study, AI-generated architectural floor plans were subjected to a qualitative assessment within a professional architectural framework. The evaluation was structured around four main criteria. First, spatial organization was examined in terms of how effectively the plans addressed user needs, with particular attention to the placement of rooms, proportionality of spatial dimensions, and the logic of functional relationships (Li et al., Reference Li, Zhang, Du, Zhang and Xie2024). Second, functionality and circulation were assessed by analyzing the adequacy of usable spaces, the clarity of circulation routes, and the overall legibility of the plan. The third criterion focused on climate responsiveness, evaluating whether the design allowed for natural ventilation through appropriate openings and whether climate-specific features, such as wide eaves in tropical regions or compact layouts in polar zones, were evident. Lastly, architectural coherence was considered, emphasizing the presence of a consistent design logic and the architectural feasibility of the proposed floor plan. This multilayered evaluation framework (Table 13) provides a critical foundation for interrogating the architectural validity of AI-driven design outputs. Although the primary focus of this study was climate adaptation, a supplementary morphological reading was conducted to examine culturally resonant architectural cues. In this context, features such as the raised platform and shaded veranda in Jakarta, the inner courtyard in Alice Springs, and the use of timber-heavy construction in Tromsø can be interpreted as echoes of vernacular typologies, albeit emergent through AI-generated outputs. While these elements may not indicate conscious cultural encoding, they suggest latent representational patterns worthy of further exploration.

Table 13. Multilayered evaluation framework

Architectural coherence

From a standpoint of architectural coherence, ChatGPT-generated plans exhibited a higher degree of internal consistency. The Madrid example, in particular, reflected a rational structural system, coherent material logic (e.g., masonry and timber use), and a buildable configuration grounded in local construction practices. These qualities suggest an implicit awareness of tectonic and spatial discipline within the AI’s generative logic.

By comparison, Microsoft Copilot’s plans, although aesthetically engaging, lacked structural legibility. In several cases, wall continuity and the implied load-bearing structure were either unresolved or implausible, undermining the technical feasibility of the architectural proposals. As such, the Copilot outputs appeared more representational than architecturally grounded.

In light of the visualizations presented in Table 2, ChatGPT emerged as the more competent model in terms of spatial articulation, functional hierarchy, climate responsiveness, and architectural buildability. Although Microsoft Copilot succeeded in producing compositionally coherent images, its outputs lacked the structural clarity and environmental sensitivity required for viable architectural design. LookX AI, which was excluded from further analysis due to fundamental representational deficiencies (e.g., absence of fenestration), further underscored the necessity of rigorous architectural criteria when assessing AI-generated content. These findings collectively suggest that while T2I models can support schematic visualization, their capacity to internalize and apply architectural principles – particularly in the context of climate-responsive housing – remains limited and warrants further refinement.

The architectural evaluation was conducted by the lead researcher, who is a licensed architect and associate professor with expertise in design theory and AI-driven generative methods. The assessment followed a structured content analysis protocol based on established architectural criteria (Li et al., Reference Li, Zhang, Du, Zhang and Xie2024), ensuring consistency and interpretive depth across both AI models. Four core dimensions – spatial organization, functionality and circulation, climate responsiveness, and architectural coherence – were used to comparatively assess the plans produced by ChatGPT and Microsoft Copilot. Each criterion was rated on a five-point scale, where higher scores indicate greater alignment with architectural logic and climate-adaptive design principles. The resulting scores are presented in Table 14 as a synthesized comparison of the two models’ architectural performance. Although the architectural evaluation involved qualitative judgments, the use of predefined criteria, a calibrated scoring system, and a structured comparison framework ensured a high degree of methodological objectivity. The evaluation did not rely on subjective preferences but rather on professional architectural standards and simulation-informed assessments, allowing for replicable and critically grounded interpretations.

Table 14. Comparative scoring of AI models based on architectural evaluation criteria

Scoring legend (1–5 scale):

1 = Very poor: No architectural value or responsiveness.

2 = Poor: Inadequate spatial or climatic logic.

3 = Fair: Acceptable but limited architectural consistency.

4 = Good: Functionally and environmentally thoughtful.

5 = Excellent: Highly coherent, climate-adaptive, and buildable.

In addition to the unweighted qualitative scores (Table 14), a supplementary weighted scoring matrix (Table 15) was introduced to enhance transparency and quantifiability. Criterion-specific weights were assigned based on their relevance to spatial performance, and final scores were computed as weighted sums. This method offers a more structured comparative lens, reaffirming ChatGPT’s relative strength in spatial reasoning while addressing reviewer concerns regarding subjectivity.

Table 15. Weighted architectural evaluation matrix

Daylight performance analysis

To assess the daylighting performance of AI-generated architectural plans, simulations were conducted using VDV on four key solar dates {spring equinox [March 21], summer solstice [June 21], autumn equinox [September 21], and winter solstice [December 21]} at 12:00 h local solar time. Each plan was evaluated across three primary functional zones: the living room, the kitchen, and the bedroom. Average illuminance values (lux) were extracted using false color mapping, providing a visual and quantitative understanding of spatial daylight distribution. The performance metric was average illuminance (lux) across these rooms during daytime hours (12:00 h). Results were compared to daylighting benchmarks derived from literature. According to Reinhart and Walkenhorst (Reference Reinhart and Walkenhorst2001) and IES Lighting Handbook (2011), the thresholds presented in Table 16 were applied.

Table 16. Daylight thresholds

In tropical and desert climates (Jakarta and Alice Springs), ChatGPT-generated plans generally met or exceeded the recommended daylight thresholds for all rooms, reflecting a relatively robust spatial openness and appropriate window placement. Particularly, the Jakarta plan exhibited an average illuminance of over 900 lux in the living area across equinox dates, suggesting potential for glare and over-illumination, yet confirming ample daylight availability. In contrast, Copilot outputs in the same locations showed underperformance in the living room (e.g., 306.1 lux in Alice Springs), but surprisingly high kitchen values (e.g., 1347.8 lux in Jakarta), possibly due to exaggerated opening placements or unshaded orientations.

In temperate and cold continental climates (Madrid and Winnipeg), ChatGPT plans performed reliably within the daylight thresholds, although solstice data revealed seasonal drops, especially in the winter months (e.g., kitchen illuminance of 193.5 lux in Winnipeg). Microsoft Copilot’s outputs in Winnipeg revealed very high illuminance levels in summer (e.g., 1740.6 lux in the living room), indicating limited passive shading logic and poor modulation of seasonal exposure.

In Tromsø’s polar context, ChatGPT’s floor plan exhibited critically low daylight performance in winter (e.g., only 9.6 lux in the living room), aligning with extreme daylight limitations at high latitudes. Even during the equinox, results hovered below minimum standards, suggesting the need for active daylighting systems or architectural adaptations like larger glazed surfaces or light-reflecting interiors.

Comparative analysis across all plans highlights that while some AI-generated layouts coincidentally met daylighting standards, these successes were not consistently climate-responsive or seasonally calibrated. In particular, Copilot’s outputs showed erratic performance patterns, with certain spaces far exceeding thresholds while others failed to meet minimum requirements. This inconsistency suggests that generative AI tools, despite interpreting prompt keywords like “natural light” or “cross-ventilation,” do not yet possess a functional understanding of solar geometry or daylight performance across seasons.

Figures 1–3 present a comparative analysis of average daylight illuminance levels across three critical residential spaces (living room, kitchen, and bedroom) based on AI-generated plans from ChatGPT and Microsoft Copilot. Overall, ChatGPT-generated layouts exhibited more consistent performance within the recommended daylight thresholds defined by Illuminating Engineering Society (2011), particularly in temperate and tropical climates. In kitchens, both AI tools produced sufficiently high illuminance values, often exceeding the 500 lux threshold; however, Copilot outputs occasionally demonstrated overexposure, suggesting a lack of refined daylight modulation. For bedrooms, where the target range is lower (200–300 lux), most AI outputs, especially from ChatGPT, surpassed minimum requirements but also risked excessive lighting in equatorial and desert contexts. The graphical comparisons underscore the variability in daylight response between models and highlight the need for postgenerative calibration when employing AI-driven plans in performance-sensitive design workflows.

Figure 1. Living room daylight performance by city and AI model.

Figure 2. Kitchen daylight performance by city and AI model.

Overall, ChatGPT-based plans demonstrated greater coherence in spatial logic and lighting performance, although even these required careful adjustment to mitigate overexposure or insufficient daylight during certain periods. The findings emphasize that current T2I AI tools can generate plausible architectural layouts, but their capacity to produce climate- and daylight-aware design solutions remains limited without postgenerative environmental refinement.

A synthesis of the findings (Table 17 and Figure 4) reveals that ChatGPT demonstrates a marked superiority in architectural criteria, particularly in terms of Architectural Coherence and Spatial Organization. However, its performance in daylight compliance remains limited, indicating a strong capacity for formal composition but a notable deficiency in meeting illumination thresholds. In contrast, Microsoft Copilot exhibits weaker architectural performance, yet achieves comparatively higher scores in Daylight Compliance, particularly in certain climate-space configurations. The radar chart (Figure 4) visually reinforces these observations, highlighting ChatGPT’s architectural advantage while simultaneously underscoring the inconsistency of both models in climate-responsive daylight performance. This suggests that while generative models excel in visual form-making, their output remains insufficiently informed by performance-oriented design logic.

Table 17. Summary comparative chart of architectural and environmental performance

^a The number of spaces within the recommended daylight thresholds was proportionally scaled from 0 to 1, and then linearly converted to a 0–5 range.

Figure 3. Bedroom daylight performance by city and AI model.

Figure 4. Summary comparative chart of architectural and environmental findings.