Visual representation in creative design

Design – whether in engineering, industrial and product design, graphic design, or architecture – relies heavily on visual outputs. These outputs, often referred to as visual representations, may be textual, static, or dynamic (Pei et al., Reference Pei, Campbell and Evans2011; Berni et al., Reference Berni, Maccioni and Borgianni2020). Designers frequently use such representations to communicate ideas and present concepts effectively. Among these, visual static representations play a critical role, serving as stimuli that evoke emotions, influence decision-making, and shape thought processes (Borgianni et al., Reference Borgianni, Maccioni, Fiorineschi and Rotini2020). Using visual stimuli is particularly significant during the early ideation stages of the design process, where creativity and exploration are paramount.

In design creativity, visual stimuli play a critical role, serving as both a source of inspiration and a potential limitation.

Borgianni et al. (Reference Borgianni, Maccioni, Fiorineschi and Rotini2020) underscored the importance of external sources of inspiration, particularly incorporating both text-based and graphic forms of stimuli, noting that intentionally selected external inputs can significantly enrich designers’ creative processes.

Agogué et al. (Reference Agogué, Kazakçi, Hatchuel, Le Masson, Weil, Poirel and Cassotti2014) confirm that the type of stimuli provided plays a critical role in shaping creative idea generation during problem-solving tasks. The scholars investigated the impact of text-based stimuli on creative flow, specifically how such examples can lead to design fixation. They identify two types of examples: “restrictive examples,” which limit possibilities by activating common knowledge and thereby reducing originality, and “expansive examples,” which introduce unexpected ideas, fostering novel solutions and enhancing originality.

Similarly, research by Cardoso and Badke-Schaub (Reference Cardoso and Badke-Schaub2011) highlights the dual nature of visual inputs: while they can spark innovative ideas, they also risk inducing design fixation, where designers become overly focused on specific attributes of a presented visual, reducing the diversity of their outputs. This phenomenon was particularly evident in their study involving industrial design students, where participants exposed to certain pictorial representations showed higher levels of repetitive patterns compared to those in a control group. Such findings emphasize the need for varied and carefully curated visual stimuli to mitigate fixation and expand creative exploration. Based on the contributions above, visual stimuli, such as images, emerge to be particularly effective due to their ability to evoke emotions and facilitate connections that might not occur otherwise. The scholars suggest that diverse visual prompts can encourage lateral thinking, enabling designers to explore unconventional ideas. In recent years, while traditional visual stimuli have long played a pivotal role in design, the emergence of AI – particularly generative AI – has revolutionized their creation and application. This transformation presents new opportunities and challenges that remain to be fully explored, including whether AI can mitigate or exacerbate fixation compared to traditional stimuli, and whether AI models themselves may be susceptible to fostering creative fixation.

Generating AI for creative design: opportunities and challenges

Artificial Intelligence (AI) has experienced rapid growth, revolutionizing traditional workflows across many fields and industries (Bordas et al., Reference Bordas, Le Masson, Thomas and Weil2024). Among the different AI branches, generative AI has gained prominence due to its ability to produce diverse and high-quality outputs such as text, images, videos, and even 3D designs (Kılınç and Keçecioğlu, Reference Kılınç and Keçecioğlu2024). Unlike AIs designed for prediction or classification tasks, generative AI excels in creating novel content by learning from patterns in large datasets (Bordas et al., Reference Bordas, Le Masson, Thomas and Weil2024; Kılınç and Keçecioğlu, Reference Kılınç and Keçecioğlu2024). Its versatility is reflected in the variety of input–output that ranges from text-to-text, text-to-code to text-to-images etcetera, expanding AI application from coding, writing to generating images and multimedia outputs (Kılınç and Keçecioğlu, Reference Kılınç and Keçecioğlu2024). Tools like DALL-E, MidJourney, and Stable Diffusion leverage these models to produce visually realistic works, potentially fostering creativity and iterative design processes in fields such as art, engineering, and architecture Hutson & Cotroneo (Reference Hutson and Cotroneo2023). Furthermore, generative AI’s capability to rapidly generate content, whether in written or visual forms, has made it an invaluable asset for both creators and industries. However, the limits and potential of generative AI are still debated; thus, further investigation is needed (Bordas et al., Reference Bordas, Le Masson, Thomas and Weil2024). The adaptability of generative AI is particularly evident in its application to text-based and image-based creative tasks, each with distinct capabilities and challenges.

Gaps and rationale of the study

Although prior studies have explored AI-assisted creativity, creative ideation, and human–AI co-creation workflows, comparatively fewer studies have conducted controlled empirical comparisons between image-generative AI systems and human designers specifically within conceptual engineering design tasks, and particularly during early-stage ideation. Moreover, existing comparative studies are frequently situated in graphic design, artistic production, or co-creative workflows, leaving conceptual engineering design comparatively underexplored. On the field of engineering design, this gap raises critical questions about whether AI can rival or surpass human creativity or whether it primarily serves as a complementary tool. Therefore, the gap is not the absence of research on AI and creativity per se, but rather the limited number of studies examining the independent creative potential of image-generative AI systems in engineering-oriented conceptual design contexts.

To address this gap, the study formulates two research questions:

• • RQ1: How do image-generative AI outputs differ from human design outputs in terms of creativity?

• • RQ2: How do different image-generative AI systems vary in their ability to improve creativity, and which creativity dimensions differ the most across AI systems?

Understanding the degree of creativity demonstrated by image-generative AI in producing design concepts is essential for evaluating its role in creative workflows. RQ1 focuses on comparing AI-generated and human-generated outputs in identifying strengths and weaknesses of AI tools in specific creativity metrics, such as novelty, quantity, and variety (Shah et al., Reference Shah, Smith and Vargas-Hernandez2003).

Additionally, while tools like DALL-E and Stable Diffusion share similar foundational models, differences in their training processes result in varying output styles, specifications, and levels of realism. These differences may impact their creative capabilities, leading to the formulation of RQ2. These open questions underscore the need for an in-depth investigation into the creative potential of image-generative AI systems, particularly in direct comparison with human designers. This study aims to address these gaps by examining the nature of AI creativity and proposing a framework for evaluating the quality and originality of AI-generated concepts.

To address these research questions, the authors conducted an experiment in which both humans and image-generative AI tools were given a design task where they had to generate ideas for a specific design problem. Human participants were instructed to produce as many ideas as possible, with a time limit of 20 minutes. The creativity of the resulting visual outputs was evaluated using the four parameters proposed by Shah et al. (Reference Shah, Smith and Vargas-Hernandez2003): quantity, variety, novelty, and quality. Based on these metrics, creativity was compared across human outputs and generative AI ones. Subsequently, creativity was also analyzed across the different generative AI tools used in the experiment (see section “Preparation of the experiment”).

By answering these research questions, this work seeks to contribute to the understanding of the current capabilities and limitations of image-generative AI systems in conceptual engineering design and to provide insights for the development of more effective human–AI design workflows. The contribution of this study lies in three main aspects. First, it provides a controlled comparison between human-generated and AI-generated conceptual design outputs within the same engineering design task. Second, it compares multiple image-generative AI systems rather than focusing on a single platform or model family. Third, it applies established engineering design creativity metrics (Quantity, Variety, Novelty, and Quality) to systematically evaluate visual concept generation by image-generative AI systems. In doing so, the study contributes to the ongoing discussion regarding the role of generative AI as an ideation support tool in conceptual engineering design.

Preparation of the experiment

In order to address the research questions, an experiment was planned in two phases. In the first phase, design students would solve a given creative product design problem. In the second phase, AI image generation programs would be asked to solve the same statement. Subsequently, different creativity parameters would be analyzed and compared, with the aim of discerning the specific areas where AI demonstrates the greatest potential to assist designers.

The experiment enlisted the participation of 20 volunteers to form a convenience sample: 10 female and 10 male, with an average age of 23.2 years (standard deviation: 1.54). All participants were final-year students pursuing the Engineering Design degree. This research adhered to the American Psychological Association Code of Ethics and received approval from the Institutional Review Board at the Universitat Jaume I (CEISH/40/2022). Each participant provided informed consent. Design engineering students were selected because the study focuses specifically on conceptual engineering design tasks. Participants therefore required a minimum level of design knowledge. On the other hand, professional designers may exhibit more advanced expertise and potentially different creative strategies, so the use of students enabled the establishment of a consistent baseline appropriate for this exploratory study.

The selected statement for the experiment involved generating the maximum number of ideas that address the following problem: “Develop a new concept of a lunch box for university use.” This specific problem was chosen because it appears in literature in studies of a similar nature (Chulvi et al., Reference Chulvi, Royo, Agost, García-García and Ruiz-Pastor2022). The solutions should be graphically represented, with each proposed alternative on separate sheets of paper. The allotted time for idea generation was set at 20 minutes, in line with the findings of Howard et al. (Reference Howard, Culley, Dekoninck, Bergendahl, Grimheden, Leifer, Skogstad and Lindemann2009), which suggest that idea generation capacity significantly declines after 20 minutes. Additionally, this time constraint has been minimized, considering that the idea generation time for AI tools will always be shorter.

To select drawing AI tools with sufficient potential for creating creative product designs, a preliminary analysis of 43 different image generation AIs was conducted. For screening purposes, the following key characteristics were considered:

• • The ability to generate distinct images even when the statement is repeated.

• • Capability to present images in a rendering style similar to the drawings that a design engineering student might produce.

• • Accessibility features.

• • No additional cost with no loss of technical capabilities. The capacity to generate generic designs.

• • Expert opinions regarding the technology they employ.

Ultimately, the chosen AI-based image-generation tools, labeled AI1-AI6 for convenience, were as follows (Table 1):

Table 1.

AI tools used in the experience

Table 1. Long description

The table consists of five columns and six data rows.

* Row 1: A I 1. Tool: Bing image creator. Underlying model family: DALL E 3. Main feature: OpenAI proprietary model. Website: w w w dot bing dot com forward slash create.

* Row 2: A I 2. Tool: Leonardo A I. Underlying model family: Stable Diffusion derivatives plus proprietary finetunes. Main feature: Artistic optimization. Website: leonardo dot ai.

* Row 3: A I 3. Tool: Playground A I. Underlying model family: S D X L. Main feature: Diffusion-based. Website: playgroundai dot com.

* Row 4: A I 4. Tool: Recraft A I. Underlying model family: Proprietary Diffusion-like architecture. Main feature: Vector-oriented features. Website: recraft dot ai.

* Row 5: A I 5. Tool: SeaArt. Underlying model family: Stable Diffusion derivatives. Main feature: Anime or artistic specialization. Website: w w w dot seaart dot ai.

* Row 6: A I 6. Tool: Stablecog. Underlying model family: Stable Diffusion. Main feature: Open-source diffusion-based. Website: stablecog dot com.

It is important to distinguish between the tool/platform level and the underlying model level. While six publicly accessible image-generation tools were selected for the study, these tools do not necessarily represent six entirely independent generative AI architectures. Some platforms operate using related diffusion-model families or derivative implementations. Therefore, the comparison conducted in this study primarily concerns the creative performance of accessible image-generative AI tools within realistic design-use scenarios rather than a strict benchmark comparison between fully independent foundational models.

All the AIs listed above are text-to-image generation tools, (image-generative AIs.) Their primary function is to create images based on textual descriptions as prompts. These tools are accessible online, making them available to a wide non-technical audience (Turchi et al., Reference Turchi, Carta, Ambrosini, Malizia, Spano, Schmidt, Santoro and Stumpf2023). The platforms where these tools are available are user-friendly, with intuitive interfaces that allow users to quickly learn and use them. No coding or model training is required, as the underlying machine learning models are already trained on vast datasets of images and text. This enables the models to understand the meaning of a prompt and generate corresponding images.

While all these tools share a common goal, they differ in the specific models and techniques they employ.

• • DALL-E 3 excels at generating highly detailed and realistic images.

• • Leonardo AI specializes in artistic styles and techniques, allowing users to create images in various artistic styles.

• • SeaArt AI is focused on artistic expression, generating images with a strong artistic flair.

• • Playground AI offers a user-friendly interface and a wide range of styles, making it accessible to users of all levels.

• • Recraft AI provides tools for editing and refining generated images.

• • Stable Diffusion is known for its versatility, capable of generating images in a wide range of styles, from realistic to abstract.

For the purposes of this study, only the core text-to-image generation capability of each tool was considered. Although some platforms may provide additional functionalities, such as image editing, video generation, or advanced paid features, these options were not used in the experiment and were not part of the selection criteria. Therefore, the comparison focuses exclusively on their ability to generate visual product concepts from textual prompts under default settings. The accessibility criterion referred to the possibility of using the text-to-image generation function without requiring paid features that would modify or enhance the core output capability. In this sense, “free access” was not considered as a technical characteristic of the AI model, but as a practical requirement to ensure that the selected tools could be used under comparable and reproducible conditions.

Development of the experiment

The experiment was conducted in two planned phases (Fig. 1). The first phase involved Design Engineering students, while the second phase artificial intelligence (AI) was involved.

Figure 1.

Methodology process.

In the initial experimental phase, voluntary participants were assembled and asked to sign the provided informed consent form. Each participant was then given drawing materials and sufficient sheets of paper. They were presented with a statement: “Develop a new concept of a lunch box to bring lunch to the university. Provide as many ideas as you can. Please draw each proposal on a separate sheet of paper.”

In the second phase, AI outputs were collected. Since the number of generated ideas may influence creativity metrics such as Variety, Novelty, and Quality, output quantity was controlled at the level of each AI tool rather than at the level of total aggregated outputs. The purpose was not to make the total number of human and AI-generated concepts identical, but to evaluate each AI tool under two human-referenced output conditions.

Two reference points were selected based on human performance:

• • Average Solutions ( $ \overline{X} $ _sol.): The mean number of ideas generated by all human participants.

• • Maximum Solutions (Max_sol.): The highest number of ideas generated by any individual participant.

These two values represent distinct but complementary human-referenced output conditions. The average reflects a typical human performance baseline, while the maximum represents the upper bound of human ideation capacity under the given conditions. These values were used as constraints in two rounds of AI prompts, each using the same design task:

• • Round 1: AI tools were prompted to generate the same number of solutions as the average human participant (coded as [ $ \overline{X} $ _sol.]S-AI_n).

• • Round 2: AI tools were prompted to generate the maximum number of solutions achieved by human participants (coded as [Max_sol.]S-AI_n).

By using both values, the study captures not only standard performance levels but also peak ideation scenarios, enabling a more comprehensive comparison between human and AI creativity. The use of two quantity conditions allows us to assess whether the relative performance of AI compared to humans is sensitive to the number of generated ideas. This is particularly relevant given that AI systems can generate outputs at a much higher rate than humans, potentially biasing creativity assessments if output quantity is not controlled. However, the comparison should be interpreted at the level of participants/tools and human-referenced output conditions, rather than as a comparison based on identical total numbers of generated ideas.

Thus, the prompts used in the different AIs were, the first one “Develop a new concept of a lunch box to bring lunch to the university,” and subsequently to $ \overline{X} $ _sol and to Max_sol. “Develop another different new concept of a lunch box to bring lunch to the university.” Default models and default styles were used for the image generation. No generated images were classified as “failed generation,” so all resulting outputs were included in the study.

Assessment method

To analyze the creativity resulting from the design process, the metric proposed by Shah et al. (Reference Shah, Smith and Vargas-Hernandez2003) was employed. This metric distinguishes four parameters: quantity, variety, novelty, and quality. The definitions of these parameters are listed below:

• Quantity: Refers to the number of solutions generated by the designer or design team. It is calculated by directly counting the outputs provided by the participants.

• Variety: Represents the degree of diversity or difference among the generated solutions. The calculation considers how each function is performed for each output, based on the level of abstraction at which differentiation occurs. To assess variety, an organizational tree (example of Function 1 of D14 in Fig. 2) is constructed:

  • • At the first level, ideas are differentiated based on the Physical Principles (PP) that satisfy the same function. PP refers to the fundamental idea or concept underlying a creative solution. It is the scientific or technical basis that justifies how the idea works. Example: In the context of a lunch box, the physical principle refers to the way food is separated, such as physically dividing it into distinct compartments.

  • • At the second level, differentiation occurs according to Working Principles (WP). WP is the practical application of the physical principle. It describes how the physical principle is implemented in a functional design or system. Example: In the lunch box, the working principle would be the specific design used to achieve physical separation, such as a single container with internal compartments or multiple containers combined into one system.

  • • The third level distinguishes ideas based on their Embodiment (E). It refers to the way the working principle translates into a tangible design. It is the physical representation of the creative concept. Example: In the lunch box, embodiment refers to how multiple containers are physically connected, such as through fixed integration or detachable, modular connections.

  • • Finally, at the fourth and last level, differences at a Detailed level (D) are considered. This level refers to the individual components that make up the design. Each part must fulfill a specific function within the system. Example In the lunch box, this includes the specific design of the connection mechanism, such as the type of locking or fitting system used to assemble and disassemble the containers.

    Figure 2.

    Example of an organizational tree used to evaluate variety (F1 of subject D14).

    

Variety is quantified using the refinement proposed by Nelson et al. (Reference Nelson, Wilson, Rosen and Yen2009), which establishes the following formula for calculating the final variety score:

$$ V=\left[10\;\left({n}_{PP}-1\right)+5\;\left({n}_{WP}-1\right)+2\;\left({n}_E-1\right)+\left({n}_D-1\right)\right]/{N}_{max} $$

where (n) represents the number of branches at the corresponding level of abstraction, and (N _max) is the maximum number of solutions.

• Novelty: Refers to the degree of originality or rarity of the generated solutions compared to existing or known solutions. It is calculated as the weighted sum of the importance of key product subfunctions, multiplied by a coefficient corresponding to the level of originality with which these subfunctions have been addressed. This coefficient takes on a value of 0 if the subfunction is not fulfilled, 3 if it has been resolved in a very common manner, 7 if it has been resolved in an infrequent manner, and 10 if it has been resolved in a completely novel way. An example is shown in Table 2.

• Quality: Represents the degree of suitability or effectiveness of the solutions generated according to problem criteria or requirements. This parameter is calculated based on the same subfunctions used in assessing novelty. However, in this case, the effectiveness of the solution provided for each subfunction is evaluated on a scale from 0 to 10. Specifically, a value of 0 corresponds to a subfunction that is not fulfilled, a value of 5 indicates basic fulfillment, and a value of 10 indicates optimal resolution. The final quality score is also quantified through the weighted sum of the values considered for each subfunction, using the weight assigned to each subfunction. An example is shown in Table 3.

  Table 2.

  Example of novelty assessment

  

  Table 2. Long description

  The table consists of five columns and two rows.

  Row 1 (Header):

  - Column 1: Level of novelty.

  - Column 2: 0 – not fulfilled.

  - Column 3: 3 – resolved in a very common manner.

  - Column 4: 7 – resolved in an infrequent manner.

  - Column 5: 10 – resolved in a completely novel way.

  Row 2 (Example of a fan – function: move the air):

  - Column 2: Does not move air.

  - Column 3: Standard fan, with blades and rotor.

  - Column 4: Uses centrifugal energy to optimize airflow.

  - Column 5: No blades, uses magnetism to move air.

  Table 3.

  Example of quality assessment

  

  Table 3. Long description

  The table consists of a header row and one data row.

  Header Row: The first column is titled Level of quality, followed by numerical columns labeled 0 through 10.

  Data Row:

  - The first column identifies the subject as Example of a fan – function: easy to use.

  - At Level 0, the quality is described as Inoperable.

  - At Level 5, the quality is described as Basic and manual controls.

  - At Level 10, the quality is described as Intelligent controls with connectivity to mobile applications.

  - Levels 1 through 4 and 6 through 9 are blank, representing the gradient between the defined anchors.

The same subfunctions and their weights considered by Chulvi et al. (Reference Chulvi, Royo, Agost, García-García and Ruiz-Pastor2022) were also utilized for the design problem used as a task in the experimental procedure. In particular, the subfunctions considered were:

• • Way of separating the food (weight: 0.6).

• • Transport and/or storage after use (weight: 0.4).

All evaluations were conducted by an expert in the field of Design Engineering, with 17 years of experience in assessing creative design. A second expert, with 10 years of experience, realized independently the same evaluations in order to analyze the agreement between evaluators. The reliability of the agreement was calculated with a Fleiss Kappa assessment, with a value of 0.952. The value has also been recalculated without considering the values of the variable “quantity,” since these values will always be coincident. In this case, the value of the Fleiss Kappa assessment is 0.857.

The study should be interpreted as an exploratory comparative design rather than as a conventional between-group experimental study. Human participants constituted a population sample with natural inter-individual variability, whereas the AI condition represented a selected set of image-generative systems evaluated under predefined generation settings. Consequently, the comparison does not aim to establish population-level equivalence between humans and AI, but rather to explore differences in creative performance under controlled conceptual design conditions.