Adaptive interior design method for different MBTI personality types based on generative artificial intelligence

1.1 Research background

The advancement of technology and the enhancement of human mastery over it have always been central to the development of architecture, leaving profound marks on the history of architecture with each technological revolution (Carpo & Yan, 2023). In recent years, with the rapid iterative development of data science and computer technology, technologies such as computational design, generative design, digital fabrication, augmented reality, and others have emerged like mushrooms after rain (Yuan & Zheng, 2023), bringing tremendous vitality and development opportunities to the contemporary architecture industry (Yuan et al., 2022). However, unlike previous technological changes, the current technological revolution of generative artificial intelligence seems to have transcended the traditional role of auxiliary tools, to some extent expanding the boundaries of architects’ imagination (Yuan et al., 2023).

In the current interior design practice, designers often face a difficult problem is that they fail to meet the needs of the owners in the design style, resulting in repeated adjustments in the design process. Such repetitive work consumes a lot of energy and time for designers. How to use the current artificial intelligence technology to efficiently assist designers to complete the optimization and adjustment of interior design schemes at the level of design style has become an urgent need to solve the problem. In order to further explore the possibilities and potentials of the intervention of generative artificial intelligence technology in the architectural and interior design processes today, the author takes as an example a project completed during the 2023 Digital Futures Creative Artificial Intelligence Model Workshop. The aim is to propose a research idea for a interior design style generation model tailored to the needs of different personality traits groups, and to realize it through the application of generative artificial intelligence technology.

2.1 The applications of generative AI in architectural designs and interior design

AIGC (Artificial Intelligence Generated Content) represents a novel approach to content generation utilizing artificial intelligence technology, following in the footsteps of Professionally Generated Content (PGC) and User Generated Content (UGC). In the domain of AI-generated imagery, it primarily relies on technologies such as Generative Adversarial Networks (GANs) and recent advancements in large-scale pre-trained models. Through the computation and fitting of massive datasets, computers can simulate human creativity and judgment, generating a plethora of relevant image content based on human needs and objectives, thereby establishing a paradigm dominated by regularized creation and deduction (Yuan et al., 2023).

Initially, with the rise of urban spatial big data, the difficulty of collecting related urban texture images and multi-source urban data has been reduced. Generative artificial intelligence technology has first found wide application in the field of computational urban design. Scholars like Yao Jiawei utilized GAN models for the morphological generation simulation of residential areas in Shanghai (Yao et al., 2021). Zheng Hao and others employed GAN models to fit urban crime rate distribution heatmaps with urban texture image data (He & Zheng, 2021). Zhang Yecheng and colleagues used GAN models to study the relationship between urban spatial morbidity and the distribution of urban facilities (Zhang et al., 2022).

In the field of architectural design, due to the relatively small amount and difficulty in obtaining architectural datasets, scholars commonly opt to start with readily accessible residential floor plan datasets. Representative examples include scholars like Huang Weixin, who utilized the pix2pixHD model to achieve the generation process from building outlines to architectural functional zoning and then to building floor plans (Huang & Zheng, 2018).

The methods employing GAN models and other derivative models for training impose strict requirements on dataset quantity, quality, and format. Although some scholars have applied them to architectural facade generation and other design fields, the quality of generation still falls short of the requirements of architectural design practice. With the emergence of new-generation AIGC painting platforms such as Midjourney and Stable Diffusion, this challenge has been correspondingly addressed. Compared to the complex file configurations and image generation methods of GAN models, Stable Diffusion, used in this study, is a diffusion model capable of achieving “text to image” and even “image to image” generation. It not only allows adjustments to the content of model-generated images through text prompts but also enables precise control over the contours and boundaries of generated images based on control nets, greatly reducing the threshold for designers’ usage.

Currently, some scholars have applied Stable Diffusion to tasks in the design and research of architectural interior spaces. For instance, scholars like Yige Fan proposed a generative artificial intelligence model inspired by the concept of “draw to shade” enabling residents to freely create shadow paintings. The model generates paintings with opaque and transparent blocks interweaving according to user preferences and needs, which are then used as window shading facilities to adjust the indoor lighting performance (Fan et al., 2023). Scholars like Hu Y. utilized AIGC technology to explore projecting artificially generated images and texts onto AR displays in the design space, proposing a design mindset from user functions to environmental generation (Hu et al., 2023).

Through the aforementioned studies, it is evident that the advantage of utilizing generative artificial intelligence models lies in the ability to generate personalized and customized design solutions based on clients’ preferences and needs. By analyzing clients’ preferences, lifestyle habits, and spatial requirements, generative models can produce design solutions that match clients’ personalities, thereby enhancing design satisfaction and user experience. This viewpoint is seldom mentioned in previous studies on generative design.

2.2 Interior design status and potential demand

In previous practices of interior design, the primary process encompassed communication of requirements between designers and clients, conceptual design and scheme confirmation, detailed scheme design, and subsequent construction and acceptance stages. Among these, the phase of communicating requirements between designers and clients often plays a crucial role in the entire design process. Only when designers fully understand the clients’ design requirements can they better accomplish design tasks. However, a challenging issue frequently encountered by designers in the process of interior design practice is the difficulty in accurately understanding or capturing clients’ requirements at the level of design styles and aesthetic preferences, leading to continuous adjustments to the schemes. This problem has persisted and occurred frequently in today’s architecture industry.

Design style is a concept derived from the summarization of common visual characteristics appearing repeatedly in multiple design cases (Chan, 1994). By summarizing different design style concepts, designers can effectively understand the preference requirements for similar interior design styles. However, due to the varying personal experiences, aesthetic inclinations, cultural backgrounds, and levels of understanding of client requirements among different designers, there often exist disparities in the perception of requirements for the same design task. People often consciously judge whether a design imagery aligns with their understanding based on their past cognitive experiences, which often leads to differences of opinion between designers and clients.

To avoid such unnecessary differences of opinion and labor wastage, it is necessary for designers to proactively explore the latent aesthetic preferences of different groups in advance of scheme design, and attempt to “tailor” specialized design schemes for them. However, relying solely on designers’ perception of aesthetic inclinations for different characteristic groups is challenging to achieve, as it would require significant effort from designers to observe and analyze the behavior patterns of clients. Certainly, designers also refer to real-life images or renderings of similar interior designs provided by clients, to create interior design schemes that closely match the proposed style type (Eckert & Stacey, 2000; Goldschmidt, 1998), yet this remains an inefficient method.

To overcome the problem of differences in design style opinions, some scholars have begun to use deep learning-based image recognition methods to determine the interior design style of reference images. For instance, Chan et al. proposed a method for identifying design styles using feature matching theory (Chan, 2000), which posits that design styles can be measured according to the degree of style, thereby providing a theoretical basis for combining deep learning methods with style recognition research. Kim J et al. proposed a method that uses reference images and a deep learning model to randomly identify and append interior design style information, defining common interior design styles, preparing datasets, and using a pre-trained VGG16 model to achieve interior design style recognition. The trained model can automatically identify the design style of given interior images (Kim & Lee, 2020). Additionally, some researchers have proposed CNN-based methods to identify design-related information in images, by constructing a large-scale image database for model training, containing approximately 400 different indoor space categories (Zhou et al., 2017). However, although deep learning methods can now be applied to recognize interior design styles, how to assist computers in generating corresponding interior design style renderings for designers still requires further exploration.

Today, with the intervention of generative artificial intelligence technology, by collecting extensive data on collective design styles and aesthetic preferences of populations, there is the potential to train specialized fine-tuning models (LORA models) to address differences in design styles for different individuals, thereby realizing “tailor-made” interior design schemes for various groups and providing effective support to improve communication and coordination between designers and clients (Fig. 1). Hence, it is particularly important to identify factors that can universally reflect differences in characteristic group design styles, with the application of the MBTI personality theory being a possible starting point.

Comparison of different interior design processes

Full size image

2.3 Study on aesthetic preference of different MBTI personality types

In the past, many scholars have been engaged in the research of aesthetic preferences among different groups. Palmer et al.’s research indicates that factors such as cultural background, gender, and age significantly influence people’s color preferences (Palmer & Schloss, 2015; Lee & Lee, 2021). Additionally, numerous studies have successfully verified the potential connection between personality traits and color preferences. This phenomenon can be explained by the viewpoint that different color tendencies can evoke specific emotional experiences in individuals (Ou et al., 2004). Previous research has confirmed the correlation between individual characteristics and aesthetic preferences. Analyzing individual characteristics can help understand individual behavioral traits in specific environments and facilitate corresponding stylized designs based on various individual characteristics. However, deeply understanding the relationship between human intrinsic traits and outward aesthetic preferences is not easy. To address this issue, researchers need to explore different personality types.

MBTI is a widely used personality classification system developed based on Swiss psychoanalyst Carl Jung’s theory of psychological types, proposed by Katherine Briggs and Isabel Myers in the mid-20th century (Myers et al., 1998). According to the combinations of these four dimensions, MBTI distinguishes people into 16 different personality types. In a study on the relationship between personality types and aesthetic preferences among college students, it was found that intuitive types prefer simpler images more than sensing types, feeling types prefer traditional images more than thinking types, and sensing types prefer modern images more than judging types. This indicates a relationship between personality types and interior environmental preferences. However, it also demonstrated that regardless of the personality type of the respondents, there is an inevitable bias towards modern styles (Lee & Lee, 2021). Kopec et al. applied personality surveys and preference index evaluation systems in their study, allowing clients to self-assess their MBTI preferences and levels and comparing design preferences to understand the influence of different personality types on design preferences (Kopec, 2012). The above studies effectively uncover the potential connection between MBTI personality types and aesthetic preferences.

Hence, this study aims to build an AI model for interior design that analyzes homeowners’ architectural style preferences based on their MBTI personality types, guiding designer’ decisions in interior space design. The innovation lies in integrating MBTI personality theory with generative AI technology, offering a smarter, more efficient, and personalized design approach for the interior design sector, aiding designers in meeting diverse homeowners’ aesthetic preferences more effectively.

This study consists of three main steps: dataset acquisition and screening, LORA model training, and evaluation of generated interior design renderings (Fig. 2).

Overall research roadmap

Full size image

Specifically, in the dataset acquisition phase, we conducted surveys targeting respondents representing 16 different MBTI personality types, collecting interior design renderings and style prompts favored by individuals of different personality types to establish a dataset of interior design styles corresponding to different personality types. Simultaneously, we performed detailed feature extraction on the image datasets of different personality types and used machine learning algorithms to summarize the color and style preferences of each personality type in interior scenes, forming corresponding preference index rankings. Next, in the LORA model training phase, based on the preference index rankings of each personality type, we screened the images in the interior design style dataset and trained the LORA model based on the screened images, thus establishing LORA models corresponding to 16 personality types. Finally, in the result validation phase, we input an interior design sketch into Stable Diffusion and fine-tuned it using the corresponding LORA model for each personality type, generating 16 interior design renderings. This process aims to verify whether this method can effectively assist designers in meeting the requirements and intentions of client-side stakeholders, providing a more efficient and accurate solution for the field of interior design.

4.1 Data acquisition and filtering

During the dataset acquisition phase for interior design styles, this study surveyed individuals representing 16 MBTI personality types. Each participant submitted 3 to 6 images of preferred residential interior designs along with descriptive words characterizing the style and ambiance. This data was utilized to compile an image dataset for the 16 MBTI personality types along with a vocabulary library of prompts.

To analyze the aesthetic preferences of different MBTI personality types, the study utilized Python programs to extract the color composition and proportions from the images in the dataset (Fig. 3), thus obtaining the common color preferences for each personality type. Additionally, to determine the preferences and weights of different personality traits for interior design styles, the study selected common modern residential interior design style types such as Industrial, Japandi, Traditional, Bauhaus Style, Mediterranean Style, Midcentury, Country House, Traditional, and Boho as survey options. This allowed the study to ascertain the common interior design style preferences for each personality type.

Analysis process of image dataset corresponding to different personality types (partial)

Full size image

After acquiring the common color and style preferences for each personality type, this study ranked and obtained preference weights for the 16 different personality types. Subsequently, based on these weights, 40 relevant interior design images representing the top 3 styles and 5 main color tendencies for each MBTI personality type were selected from the acquired image dataset as the training dataset (Fig. 4). These images were then used to train specific LORA models for each personality type.

Analysis results of different personality traits regarding interior design style preferences and weights

Full size image

4.2 LORA model training

During the training process of the LORA model, this study utilized the open-source LORA model Web UI platform for relevant model training. Taking the training of the LORA model for the INFP personality type as an example, this paper provides a detailed explanation of the entire model training process. Specifically, before training, the images in the image dataset need to be labeled, meaning each image needs to be accompanied by a separate txt file describing the content of the image (including prompt words) (Fig. 5).

Method of training images and their labels

Full size image

After obtaining the tag files, the relevant training parameters were set up and the LORA model Web UI platform was imported for training. The chosen base model for training was “anything-v5-PrtRE.safetensors”, and the main parameters for model training are shown in Table 1. In particular, the learning rate adjustment strategy selected was “cosine_with_restarts” for the lr_scheduler parameter to avoid underfitting and overfitting during the model training process. The remaining parameters were kept at their default settings.

This study conducted training for the LORA models of 16 different personality types based on these parameter settings, while employing default data augmentation methods for data expansion. It should be noted that for architectural or indoor scene settings, the network_dim parameter often requires a higher value. Through multiple tests and comparisons, it was found that the image generation results were better when the value of network_dim exceeded 128. Taking the training situation of the LORA model for the INFP personality type as an example (Fig. 6), the blue curve represents the learning curve results when network_dim = 32, and the red curve represents the learning curve results when network_dim = 128. It can be significantly observed that when network_dim takes a larger value, the loss value of model training is significantly improved, indicating a better fit to the data.

LORA model training process

Full size image

In addition, analyzing the preview images generated every 10 iterations during the model training process helps to gain a more intuitive understanding of the specific training process and the generation capabilities of the model. Taking the training process of the LORA model corresponding to the INFP personality type as an example, (Fig. 7) shows the preview images generated by the LORA model every 10 iterations. As the number of iterations in model training increases, the quality of the generated images significantly improves. Around 60 to 80 iterations, the richness of image generation details visually meets the target requirements of this study. However, as the model continues training, there will be an excessive amount of unnecessary details in the generated preview images, making the generated images appear more cluttered. Therefore, the model training process does not necessarily improve with more iterations. Instead, it is essential to find the most appropriate time during the training process. In the training process of this LORA model, the author saved the model obtained after 70 iterations for subsequent generation experiments.

LORA model preview image analysis

Full size image

For the training results of LORA models corresponding to the other 15 personality types, the training methods were the same as above, and the results obtained after training were very similar to the training results of the LORA model for the INFP personality type. Ultimately, this study obtained 16 LORA models tailored to specific MBTI personality types, providing a foundation for generating interior design effect images preferred by different types of users.

5.1 Interior design effect generation

After inputting the content prompts for generation, Stable Diffusion is able to generate images that align with the design style and color preferences of the corresponding personality type by fine-tuning the LORA model specific to that type.

To validate whether the interior design style effect images generated by fine-tuning the LORA models on the Stable Diffusion platform can meet the requirements of owners with different MBTI personality types, this study adopted the following process to generate and test personalized interior design effect images for users of different MBTI types:

1. (1)

   Select a hand-drawn interior design sketch as the base image for generation. This sketch contains the basic outline of the interior space and furniture layout but does not include detailed furniture design features or color information.

2. (2)

   Import the base image into the Stable Diffusion platform and use the “image-to-image” method for image generation. The base image is simultaneously used as a Control Net for image edge and composition control, and the Canny model is used for image edge detection.

3. (3)

   16 LORA models of MBTI type trained before were respectively loaded, and the same Positive Prompt and Negative prompt were used to generate images, and then 16 interior design renderings based on the corresponding preferences of different MBTI personality types were generated.

In total, this study generated 16 interior effect images tailored to different MBTI personality types (Fig. 8). The generated images retain the spatial structure of the original sketch while presenting the color styles, decorative elements, and atmospheric characteristics of different personality types, achieving the goal of creating interior effect images with strong specificity.

Interior design renderings of the same draft after fine tuning of different LORA models

Full size image

5.2 Analysis of interviewee satisfaction

To evaluate the satisfaction level of individuals with different MBTI personality types towards the interior effect images generated through specific LORA fine-tuning, this study first selected 210 respondents for the test. These respondents all clearly identified their MBTI personality types and ranged in age from 20 to 50 years. It should be noted that due to the varying rarity of different MBTI personality types in the population, there may be uneven proportions among respondents. For example, the INFJ personality had the largest number of respondents, with 21 participants, while the ISTJ and ESTJ personalities had the fewest respondents, each with only 7 participants. The distribution of other personality types was relatively even, with the number of respondents ranging from approximately 10 to 15 individuals. The survey questionnaire prepared by this study included 16 different styles of interior effect images generated by loading different LORA models using the same interior sketch base image, presented in random order. Respondents were asked to select three interior effect images they preferred from these 16 images. Subsequently, by tallying the proportion of votes for different interior effect images from respondents of different MBTI personality types, the study analyzed whether the LORA model had learned the aesthetic preferences of different MBTI personality types. After the survey was completed, a total of 203 valid questionnaires were received. Through statistical analysis of the voting preferences of each respondent for the 16 images, the following results were obtained (Fig. 9). In the figure, the x-axis represents the 16 interior style effect images generated by different LORA fine-tunings, the y-axis represents the respondents of 16 different MBTI personality types, and the numbers in the heatmap represent the number of votes each type of respondent gave to the corresponding image. The higher the number, the higher the level of preference for that image among respondents of the corresponding MBTI personality type.

Thermal map of image preference degree voting statistics

Full size image

To explore whether the interior effect images generated by fine-tuning different LORA models can truly attract respondents with the same personality type, the study counted the number of votes for the preferred interior effect images by respondents of each personality type. The study then ranked the top five preferences for each personality type based on the proportion of votes and observed the degree of preference for different LORA model fine-tuned effect images among different MBTI personality types. The results are shown in (Fig. 10).

Image preference ranking

Full size image

From the figure, it can be observed that for respondents with personality types such as INFP, INTJ, INFJ, and ESTJ, the interior effect images fine-tuned by their corresponding LORA models received the highest number of votes among their preferred images. For respondents with personality types such as ENFJ, ISFP, INTP, ESFP, ESTJ, ESTP, ENTJ, and ENFP, the interior effect images fine-tuned by their corresponding LORA models tied for the highest number of votes with other images. However, for respondents with personality types such as ISTP, ISTJ, ISFJ, and ENTP, the images with the highest number of votes were not necessarily generated by their corresponding LORA models.

Based on the research results, the following conclusions can be drawn:

1. 1.

   The aesthetic preferences of different respondents are not solely determined by their MBTI personality types but are also influenced to some extent by factors such as their family background, educational experiences, and life experiences. Therefore, relying solely on training LORA models corresponding to different personalities cannot accurately predict their aesthetic preferences. To achieve more accurate predictions of aesthetic preferences in interior effect image generation, additional factors such as style description prompts are necessary.

2. 2.

   Conversations with some respondents revealed another reason why the prediction results for different personality aesthetic preferences were not ideal. This was because the experiment used the same sketch to generate interior effect images, implying consistency in interior space layout. However, different personalities may have different evaluations of this space. Therefore, when using AIGC tools for interior effect image generation, differences in the layout preferences of different personality types need to be considered.

3. 3.

   Although only ten personality types of respondents selected interior effect images fine-tuned by their corresponding LORA models as their most preferred, a pattern was observed. For J-type personalities, their preferred effect images tended to be those fine-tuned by J-type LORA models, and the same applied to P-type personalities. This indicates that J-type personalities tend to prefer interior designs that are more regular and orderly, while P-type personalities prefer more flexible and varied interior design scenes. Therefore, in MBTI personality types, J-type and P-type personalities play a significant role in people’s aesthetic preferences, allowing designers to effectively differentiate between the preferences of homeowners based on this dimension.

6.1 Interior designer’s assistant tool

This method serves as an invaluable tool for interior designers to swiftly generate personalized interior design solutions tailored to clients’ MBTI types. Designers only need to input the client’s MBTI personality characteristics and room layout sketches. By providing relevant design prompts, they can use the Stable Diffusion platform to generate interior design renderings that align with the client’s preferences. This includes interior color schemes, furniture arrangements, and decorative elements. The method significantly reduces the time and effort interior designers spend on design iterations, lowering the cost of repeatedly modifying and rendering interior effect images. Designers can also optimize designs based on generated renderings to enhance overall interior design efficiency.

6.2 Interior decoration company user experience enhancement service

Interior decoration companies can enhance user experience services by collecting clients’ MBTI personality types and past interior design photos. This information can be used to establish a correlation database between MBTI and interior design styles. The database, in the form of user profiles, can provide designers with corresponding design suggestions and strategies. This facilitates better interior design style scheme selection. Additionally, based on the correlation database, decoration teams can optimize final designs through improved communication with clients. This data-driven design approach significantly improves user satisfaction with design schemes and the overall decoration process.

6.3 Personalized home furnishing recommendations service

Home furnishing companies can offer consumers a personalized home product configuration recommendation service based on MBTI personality type testing (Fig. 11). Users provide their MBTI personality type, and the company utilizes artificial intelligence models to offer personalized home solutions, including furniture, soft furnishings, and accessories. Simultaneously, companies can generate home product advertisements tailored to specific MBTI personality types, achieving precise targeting and significantly enhancing the consumer shopping experience.

Personalized home furnishing recommendations service

Full size image

6.4 Personalized scene mode setting for smart homes

The research outcomes can be integrated with smart home devices and systems to create personalized smart scene modes that cater to the preferences of different MBTI personality types (Fig. 12). The system, considering the user’s MBTI type and personal preferences, adjusts lighting brightness, color temperature, and color effects through intelligent home system management and control. This fulfills users’ personalized needs, providing a comprehensive personalized usage experience. This innovation propels the development of the smart home field, creating more intelligent, comfortable, and personalized living environments for users.

Customization of personalized scene modes

Full size image

6.5 User-engaged VR interior design experience service

Building on this method, a user-engaged virtual reality (VR) interior design system can be developed, allowing users to experience various interior design schemes in real-time within a virtual space (Fig. 13). Users input their MBTI personality type and relevant requirements, don VR devices, and experience diverse personalized virtual design effects. Users can interact with the system, providing feedback. The system, in turn, adjusts design schemes in real-time based on user feedback and preferences, enabling immersive assessments and selections of design schemes.

Virtual reality (VR) interior design generation system

Full size image

This research, grounded in MBTI personality theory and employing artificial intelligence technology, aims to generate interior design renderings for different MBTI personality types. Experimental results indicate that this method efficiently captures the design style preferences of users with different MBTI types and intelligently generates design renderings that meet individualized preferences. While opening new possibilities in the field of interior design, the study faces certain limitations. Challenges include expanding the dataset’s scale and diversity for better model generalization and adapting the training and optimization process to enhance image generation and user satisfaction. Furthermore, exploring the correlation between MBTI personality types and other architectural design requirements can broaden the application scope of creative artificial intelligence models. In the foreseeable future, artificial intelligence in the built environment will revolutionize how we process information, presenting opportunities and challenges for architects and the industry as a whole. AIGC technology, whether optimizing traditional design workflows, aiding designers in critical tasks, or empowering architects with new styles and thinking methods, will be a significant challenge to traditional wisdom and a catalyst for innovation.