A Study on the Applicability of Generative Artificial Intelligence in the Automotive Design Process

2.1 Reasons for Choosing DALL-E and Midjourney

DALL-E and Midjourney were selected among the various AI design tools available because they are the most widely used applications.²⁾ In fact, based on a recent survey conducted by Platform.io, 77.2 % of 500 artists with experience using text-to-image generative AI reported that these AI tools were useful in their work. Furthermore, in Platform.io’s AI preference survey, DALL-E and Midjourney ranked first and second, respectively (Fig. 1). According to research by Statista, the two tools have a combined market share of 51.15 % (Fig. 2), making them the most influential tools in the market.

Fig. 1

Surveyed Artists’ Preferred AI Tools (%)

Fig. 2

Global Market Share of the Biggest Generative AI Image Tools in 2024 (%)

2.2 Utilizing AI in the Design Process

Bang and Kim³⁾ explored how to utilize AI at all product design process stages. Specifically, they argued that utilizing generative AI during the ideation and development stages can complement designers’ creativity while saving time and resources in the product design process. Jeong and Choi⁴⁾ claimed that AI tools can be primarily utilized in tasks such as idea generation, design refinement, and prototype creation. However, they suggested that designers’ requirements for support tools will still vary depending on the design field, design output type, and company structure.

Hyundai Motor Group and Volkswagen Group are examples of actual corporate use of generative AI in the transportation design field. Both companies are customizing and using the program. Hyundai AIRS Company,⁵⁾ a subsidiary of Hyundai Motor Group,⁵⁾ developed dedicated software and implemented it in its wheel design process.

Audi, a brand under the Volkswagen Group, has applied a design tool called Felgan⁶⁾ to its wheel design process. Wheel design requires fewer considerations than the overall exterior design of a vehicle. This seems to make it a suitable candidate for piloting generative AI tools capable of generating numerous design options in a short time.

Fig. 3

Generative AI example: Audi Felgan

Fig. 4

Generative AI example: Hyundai AIRS Company

Currently, AI tools are mainly used for 2D designs, such as frontal images of wheels. However, with rapidly advancing AI technology, it is expected that soon, full 3D images of vehicles will be generated without distortion.

2.3 Analysis of Prior Research

We examined past research cases that utilized AI-based tools in the design process in the design industry. This study primarily examined papers published after 2022, when DALL-E and Midjourney launched their services, as synthesized in Table 1.

Table 1

Generative AI research

Chen et al.⁷⁾ noted that automotive designers face significant challenges in verbalizing their ideas while using AI tools. Their experiments demonstrated that mobility designers’ work integrates both verbal and nonverbal expressions, making it difficult to fully reflect their opinions when inputting prompts to the AI. Therefore, they stressed the necessity of a precise understanding of how the AI responds to specific keywords to accurately use the tool.

Hwang⁸⁾ designed a poster with a unified theme using various AI tools. While recognizing the usefulness of AI as a tool, the author emphasized the importance of designers setting specific prompts as supervisors. The researcher also pointed out the limitation that the study relied solely on qualitative analysis methods and thus could not be extended to quantitative findings.

Na⁹⁾ studied the use of Midjourney in the idea generation phase of product design. The author assessed its effectiveness by generating results from various prompts and consulting designers regarding the results. The researcher also emphasized the need to integrate the designer's experience and sensibility to overcome limitations in specific details, although it would be helpful for most initial design drafts and idea generation.

Cheng and Kim¹⁰⁾ explored the potential for creating new value using generative AI tools in spatial design. Adopting the Double Diamond model, they proposed a method for utilizing generative Stable Diffusion in each phase to perform the process more efficiently. The results confirmed that AI tools are indeed supporting designers in developing more creative and innovative designs.

On the other hand, Lee and Ko¹¹⁾ noted the paucity of research on AI image generators in architecture, so they explored their use in the architectural design process. They particularly experimented to determine whether Stable Diffusion could generate architectural concepts and three-dimensional architectural images. Although improvements were necessary to construct specific three-dimensional forms, they observed that the issues were likely to be resolved in the near future. Consequently, they anticipated that while increased productivity among architects would enhance the quality of architectural spaces, core benefits could be concentrated in the hands of a select few.

Although studies exploring AI applications in the transportation sector remain underexplored, the aforementioned relevant studies emphasize the potential of AI, as they examined cases in which AI design tools were utilized in various design fields to improve design work efficiency. However, the part where the utility was demonstrated was still limited to the early stages of the overall design process. Additionally, even though it provides invaluable insights on using AI tools, designers must have a more thorough understanding of the correct procedures to ensure that the AI functions properly.

3.1 Experimental Design

The experiment was performed in four stages. In the first stage, the same prompt was input into DALL-E and Midjourney to generate images. In the second stage, the generated images were input into each tool to extract prompts. In the third stage, emotional vocabulary related to car images was collected based on a literature review, and similar emotional vocabulary was classified into several groups. In the fourth stage, the correlation between the extracted prompts and the initially input prompts was analyzed. The similarity between the two prompts was measured based on the emotional vocabulary grouped in the previous stage, and the analysis was conducted using the Jaccard index. The overall experimental flow is illustrated in Fig. 5.

Fig. 5

Flow of experiment

The Jaccard index is a value that measures the similarity between two sets. When comparing two texts or numeric data, it expresses the degree of similarity in numbers. The formula for the Jaccard index is shown in Fig. 6. In case the original input prompt is set as A, and the image generated by the AI tool is converted back into text as set B, the Jaccard index of A and B will be calculated to measure the similarity between them. The Jaccard index provides a clear numerical value of the text similarity between two prompts, serving as a crucial indicator for relatively comparing the operational performance of the tool. It was assumed that the tool with a higher index can be used as a basis for identifying which tool better reflects the user's intent.

Fig. 6

Jaccard Index formula

While DALL-E does not allow variable manipulation when inputting prompts, Midjourney can artificially distort images using three parameters: stylization, weirdness, and variety. In the experiment, all three conditions were set to 0 to control the variables.

3.2 Overview and Collection of Emotional Vocabulary

Prior to analyzing the correlations between prompts, it is necessary to adjust the method for calculating the Jaccard index. Treating each word in the prompt as an element would lead to an excessively large number of denominators, yielding an exponent very close to zero. Therefore, articles in English sentences were excluded, and unnecessary nouns were not considered elements.

Emotional vocabulary, a key technique used to study customer emotions, is commonly used to translate consumers’ emotional needs into product design elements. This study selected vocabulary appropriate for evaluating automotive design through collecting and refining relevant literature. First, 380 words were selected from papers examining emotional vocabulary in product design by Helander et al.,¹²⁾ Chae et al.,¹³⁾ Lim and Na.¹⁴⁾ From these, duplicate expressions were removed to condense the list to 149 words. Second, 104 words suitable for describing transportation vehicle design were selected. Third, the 104 words were divided into eight groups according to their visual formative qualities. The classification was based on the survey results obtained by Helander et al.¹²⁾

Group A was characterized by a feminine feel, Group B by a compact feel, Group C by a trendy feel, Group D by a classic feel, Group E by a luxurious feel, Group F by a masculine feel, Group G by a sporty feel, and Group H by a mechanical feel. The classifications are labelled in Table 2.

Table 2

Adjective keywords grouped for automotive design

4.1 Image and Prompt Generation Results of DALL-E and Midjourney

Initially, input prompts were created for DALL-E and Midjourney, and image generation was instructed based on these prompts. For both DALL-E and Midjourney, a single prompt generated four images simultaneously. The participants selected the image that best represented the input sentence and showed minimal distortion. The selected images were then uploaded to each participant and converted back into prompts. The generated results are compared in Fig. 7.

Fig. 7

Images and Prompts by DALL-E/Midjourney

4.2 Prompt Agreement Based on Jaccard Index

The baseline input prompt was input into two AI tools, and images were generated using the text-to-image function. The images were then input into two AI tools, and the image-to-text function was used to generate prompts.

The results of replacing the three prompts with emotional vocabulary groups according to Table 2, summarized above, are represented in Figs. 8 ~ 10, respectively. Group A was assigned red, Group B orange, Group C yellow, Group D green, Group E blue, Group F purple, Group G black, and Group H gray. When substituting words, duplicate words were excluded, and words within the same group were designated by sequentially appending numbers, such as A1, A2, A3, etc.

Fig. 8

Analyzing the input prompt

Fig. 9

Analyzing the prompt converted by DALL-E

Fig. 10

Analyzing the prompt converted by midjourney

The Jaccard index was calculated by comparing the input prompts and the prompts generated by each AI tool with the original text. The Jaccard index is the intersection of two sets divided by their union, which has a value between 0 and 1. The equations for calculating the Jaccard index are displayed in Figs. 11 and 12, respectively.

Fig. 11

Jaccard index between the input prompt and the prompt extracted from DALL-E Image

Fig. 12

Jaccard index between the input prompt and the prompt extracted from the midjourney image

Based on this, DALL-E obtained a value of 0.6 when compared to the original text, and Midjourney obtained a value of 0.467 (rounded to the fourth decimal place), as specified in Table 3.

Table 3

Calculated result of jaccard similarity

4.3 Interviews with stakeholders

Interviews were conducted with 12 second-year industrial design students specializing in transportation design to analyze the experimental results from a human perspective.

Considering the difficulty of mathematically analyzing the similarity between two texts, we indirectly compared the prompts written by researchers and the prompts converted by AI, which showed that the AI prompts are closer to the images generated by generative artificial intelligence.

Before the interview, the participants were briefed on the experiment flow. They were shown input prompts written by the researcher and informed that two generative AI tools were used to generate images based on these input prompts. The names of the software used in this process were not disclosed to avoid any biases that the subjects might have based on their previous experience using the selected tools.

After showing the images generated by the first AI tool (DALL-E), the participants were asked to read the text converted back into the original text using the same AI tool. They were then instructed to evaluate which of the input prompts and the prompts extracted from DALL-E had the highest relation to the previously viewed image. In addition, using the same process, they were asked which of the texts converted by Midjourney or the original text best described the image generated by Midjourney. The results of all 12 college students’ responses are summarized in Table 4.

Table 4

Student survey results

Nine out of twelve respondents reported that the prompts re-extracted from DALL-E more accurately represented the images generated by DALL-E than the original input prompts. For the images generated by Midjourney, seven out of twelve respondents judged that the text reinterpreted by Midjourney better matched the images than the original text. When the results were presented, 75% of respondents selected the text regenerated by DALL-E, while 58% selected the text generated by Midjourney (Figures 13 and 14) (Figs. 13 and 14).

Fig. 13

Student survey results (DALL-E and input prompt)

Fig. 14

Student survey result (Midjourney and input prompt)

4.4 Results Analysis

After comparing the two results using the Jaccard index, DALL-E (0.6) exhibited a higher value than Midjourney (0.467), suggesting that DALL-E produced results that are relatively closer to user intent than Midjourney.

Student interviews demonstrated a 9:7 preference for DALL-E, confirming its superiority in expressive clarity.

In both quantitative and qualitative evaluations, DALL-E outperformed Midjourney. The Jaccard index indicated that the former had a 28.48 % higher text match rate (rounded to three decimal places) and a 28.58 % higher preference rate in the human preference survey (rounded to three decimal places) than the latter. Because the quantitative analysis (Jaccard index) and qualitative analysis (interview evaluation) aligned, DALL-E can be interpreted as better reflecting user intent as a prompt-based generative AI. Conversely, Midjourney can be interpreted as being relatively more likely to expand designers’ ideas in unexpected directions.

The experimental results of the two AIs can be summarized as follows: DALL-E can express input prompts very specifically and quantitatively accurately reflect the user's original intent. Midjourney can generate unexpected images even with somewhat vague prompts, providing visual inspiration for users in the early ideation stages. This study provides valuable insights into the detailed strengths of each program, which can contribute to improving the efficiency of visualization in the automotive design process.