A Guided Prompting Interface for Supporting Creative Intent in AI Image Generation

The formative study begins with qualitative research—including semi-structured interviews and observational sessions—with a diverse group of novice digital creatives. Participants are recruited based on criteria such as limited prior experience with generative AI in creative contexts, and demographic information (e.g., age, education, creative background) is systematically documented. The goal of this phase is to uncover pain points, desired features, and user expectations related to the prompt creation process. The collected data are analyzed using thematic analysis to identify common challenges, which subsequently inform the design of the proposed method.

In the subsequent phase, a mixed-methods evaluation is conducted to compare the proposed method with a baseline text-based prompt system. A broader sample of participants, representing varied creative and demographic backgrounds, is recruited. Participants complete creative tasks using both systems, and quantitative measures (e.g., task completion time, creativity ratings, satisfaction scores) are collected alongside qualitative feedback through follow-up interviews. Statistical analyses are performed to compare performance across conditions, while qualitative data are coded to extract insights regarding usability and creative impact.

A total of 16 participants (8 males, 8 females) were recruited, all of whom had interacted with AI image generators between one and five times (Table 1). Participants with zero experience were excluded to ensure they could meaningfully explore the tools while still exhibiting novice behaviors. This recruitment criterion drew on Pirolli and Card’s (1999) Information Foraging Theory, which indicates that users with fewer than five interactions typically engage in exploratory behaviors rather than forming systematic strategies. Likewise, Luger and Sellen (2016) highlight that users with limited AI experience often encounter a gap between expectations and real-world performance, underscoring the need for structured guidance. Participants were recruited using a snowball sampling approach, beginning with individuals who met the selection criteria, followed by recruitment through social networking services and graduate school communities. Participation was entirely voluntary, and eligible participants were selected from those who expressed interest. The study was conducted between October 1 and October 31, 2023. Each session lasted approximately 50 minutes and included both experimental tasks and a semi-structured interview. All sessions were conducted individually and in person, at a time and location convenient for the participant.

Table 1

Participant demographics and prior experience with AI image generators

Each participant began by exploring AI image generation tools—Midjourney, Playground, and DALL·E—along with community resources like the Creation feed (a repository of user-generated prompts, parameters, and images). Participants were then guided through five observational tasks (Table 2), each designed to evaluate how they composed prompts and adapted their strategies under different interface constraints. Following these tasks, participants took part in in-depth interviews, focusing on their initial goals, alignment with expectations, and any challenges encountered.

Table 2

Overview of observational tasks

To assess prompt writing abilities and identify areas for improvement, the five tasks include:

• 1. Generate an image closely matching a reference provided by the researcher using only text prompts.
• 2. Transform the generated image into a Van Gogh–inspired style by modifying the prompt.
• 3. Experiment with style filters to produce a personally desired image, gauging the impact of pre-defined visual options on creative outcomes.
• 4. Reference prompts and images in the Creation feed to replicate or adapt an appealing style.
• 5. Leverage automated prompt suggestions to refine an existing prompt and observe how these suggestions shape creative decisions.

Upon completing the tasks, each participant took part in a 20-minute interview. These sessions provided qualitative insights into user behaviors, pain points, and contextual factors influencing prompt creation. A thematic analysis of the interview transcripts was conducted to distill user needs, which subsequently guided the design of the proposed method’s step-by-step interface.

① Difficulty in Writing Specific Prompts

Participants frequently found it challenging to determine the optimal length, wording, and technical details required for crafting effective prompts. They also encountered difficulty navigating platform-specific commands and features.

“I had to use Discord commands (for Midjourney), but I didn’t know which ones were available. It was difficult to use the prompts correctly.” (P2)
“I don’t know how long or detailed my prompt should be or which elements I should include for producing a good image.” (P6)

② Difficulty in expressing abstract ideas

Translating abstract concepts or mental images into precise text descriptions proved especially difficult. Expressing mood or style nuances often resulted in vague or overly general prompts.

“When I had to write the prompt myself, it was hard to convert the image in my mind into words.” (P10)
“I thought I had described the image in detail, but I wasn’t precise enough. Choosing the right words for style, composition, and mood was very difficult.” (P9)

③ Language barriers and nuance conveyance

Non-native English speakers noted that subtleties in their intended meaning were sometimes lost during translation or rephrasing, leading to unintentional mismatches between their ideas and the generated outputs.

“The nuance of the words I wanted to express changed during the translation process from Korean to English.” (P3)
“My idea of ‘concept art’ is a simple sketch focusing on the character, but the AI understood it as a more refined image. It handles complex images well, but struggled to produce rough, simpler sketches.” (P15)

① Filter utilization

Participants generally appreciated the straightforward style filter approach in Playground. By selecting pre-defined filters with sample images, they could bypass extensive text-based prompt crafting.

“I liked not having to write out the prompt directly since the expected style of the image was generated through the filter.” (P10)

However, the limited variety and ambiguous naming of filters sometimes hindered the user experience, restricting the range of creative possibilities.

“There weren’t many filters, so it was limiting. Also, some filter names were confusing. For example, I tried ‘Warm box’ but had no idea what style that would produce.” (P8)

② Creation feed utilization

Having access to verified prompts and corresponding images helped participants quickly replicate or adapt styles without the need for extensive trial and error.

“By referencing the prompts of well-made images, I was able to create the desired images quickly and easily. I think it would have been difficult to create them without referring to the prompts in the creation feed.” (P4)

③ Prompt Recommendation

Users found AI-generated suggestions, especially via ChatGPT in DALL·E, beneficial when they lacked ideas or struggled with linguistic nuances. The tool often refined their rough concepts into more detailed and coherent prompts.

“I sometimes struggle with English nuances, but the AI was helping me with that difficulty. It felt like the AI was refining what I wanted to express in more detail.” (P5)

However, some participants also mentioned the risk of overly verbose suggestions that diluted the main creative intention.

“The GPT-generated sentences were quite long, which made it hard to grasp the main idea. I had to revise them to capture my core intention.” (P10)

Overall, these findings underscore the tension between ease of use and creative control. While filter-based or AI-driven aids can streamline prompt creation, users also desire more transparency and personalization to effectively realize their artistic visions.

While prior text-to-image services (e.g., Midjourney, Playground, DALL·E) rely heavily on free-form prompts, many beginners struggled with prompt composition and structure. This gap often led to suboptimal results, undermining user confidence. Rather than merely fixing usability issues, the step-by-step approach adopted in this study seeks to democratize creative expression by scaffolding each phase of prompt writing. Research in user experience design shows that structured guidance can mitigate complexity and uncertainty in sophisticated systems (Grossman et al., 2009). Applying these insights to a creative context enables novices to iteratively refine their ideas, effectively bridging the gap between conceptual imagination and final output. Future investigations incorporating open-ended tasks may further expand users’ sense of creative freedom by allowing exploration of personally meaningful artistic visions rather than reproducing a single reference image.

Participants reported that stylistic or thematic exploration became easier when visual cues or filters were available. Although Playground’s single-filter approach reduced entry barriers, users expressed a desire for greater variety and clearer filter naming. A multi-filter selection mechanism therefore holds promise as a method for guiding novices in combining artistic styles, effectively bridging purely textual prompting with more visually oriented workflows.

In addition, community feeds were highly valued, providing real-world examples of creative outputs. Within the context of digital creativity, such community-generated prompts can inspire new directions and reduce the intimidation often experienced during early creative exploration. By integrating these features into the proposed method, the goal extends beyond streamlining functionality toward broadening users’ creative potential.

Findings indicate that AI-driven prompt recommendations (e.g., via ChatGPT) do more than resolve linguistic challenges: they can stimulate new creative ideas and encourage users to expand beyond their initial concepts. Non-native English speakers, in particular, noted the value of nuanced vocabulary suggestions that helped refine their artistic goals. However, the risk of overly verbose or tangential AI-generated prompts highlights the need for focused guidance that preserves the user’s core creative intent. This functionality demonstrates that AI assistance can actively contribute to the creative process, enhancing expressive capacity while maintaining alignment with the user’s intended vision.

Taken together, these strategies move beyond simple usability fixes. By supporting creative freedom, incorporating community-based inspiration, and providing adaptive AI assistance, the proposed method seeks to foster deeper engagement with digital art-making processes. Future design refinements will incorporate more open-ended tasks, expanded measures of creative outcomes, and culturally diverse participant samples to capture the full scope of artistic exploration. This approach demonstrates how a structured, guided interface can empower beginners to explore new artistic possibilities and develop unique creative expressions that integrate design, technology, and creative practice.

The top-down research phase provided an initial design rationale by identifying commonly accepted components of effective prompt construction across academic studies and practitioner-oriented guidebooks. These sources consistently emphasized dimensions such as subject definition, stylistic framing, compositional control, emotional tone, and lighting conditions.

However, rather than directly transferring these components into the interface, they were selectively adapted to align with novice users’ mental models identified in the formative study. This ensured that theoretically grounded prompt elements were translated into interaction-friendly categories suitable for a step-by-step design flow.

Table 3

Six high-level categories identified through top-down research

The bottom-up analysis of approximately 100,000 user-generated prompts served as a validation layer for the proposed interface design. By identifying frequently used keywords and prompt structures within Midjourney’s Creation feed, this analysis confirmed that the selected categories and filters reflect actual user practices rather than designer assumptions.

Importantly, this data-driven process informed not only category selection but also filter granularity, ensuring that the implemented UI supports both common and expressive creative intents observed in real-world usage.

Table 4

Sample extracted keywords and their usage frequencies from user-generated prompts, illustrating the bottomup validation of our six categories

The synthesis phase culminated in the implementation of an experimental service interface that operationalizes the six-category framework through 65 visual filters. This interface represents the primary experimental artifact evaluated in the subsequent usability study. By enabling users to construct prompts through interaction with visual and categorical elements—rather than relying solely on textual formulation—the design directly responds to formative findings regarding cognitive load and uncertainty. As such, the implemented interface functions not only as a tool for image generation, but as a research probe for examining how structured design interventions influence creative interaction with generative AI systems.

Table 5

Final Categories and filters of the Proposed Method

The navigation interface is organized around six core categories—Subject, Style, Artistic Reference, Composition, Mood, and Tone. Each category reflects key components identified through both top-down analysis and bottom-up user data, aligning theoretical models of prompt construction with actual user practices. This step-by-step structure is intended to reduce the cognitive load often associated with prompt creation, particularly for novices who may struggle to translate abstract ideas into precise textual descriptions.

Users may choose to skip categories that are not relevant to their creative objectives, thereby retaining flexibility to focus on elements most aligned with their intended vision. Figure 1 illustrates the category selection screen, where each category is accompanied by a concise explanation to clarify its potential influence on the generated image. By structuring the interface in this manner, the proposed method directly addresses the need for guided prompt construction, a major pain point identified among beginners during the formative study.

Figure 1

Category selection screen. Each category is briefly explained to guide novice users through essential aspects of prompt creation, helping them articulate subject matter, style, and other creative dimensions in a structured manner

To further support users who may be uncertain about how to describe specific visual attributes, each category provides up to 15 filters, resulting in a total of 65 options across all six categories. Each filter is presented with a representative image to offer immediate visual feedback—an approach grounded in formative study findings indicating that novices respond more effectively to concrete visual examples than to abstract textual descriptors. Users can apply multiple filters within and across categories, allowing combinations of stylistic, compositional, and emotional cues to generate more nuanced and intentional prompts.

The selected filters are dynamically reflected in the prompt window, eliminating the need for users to manually articulate or recall specialized terminology. Once selections are finalized and the “Create” button is activated, the system produces four high-quality image outputs that can be viewed, refined, or downloaded as needed Figure 2 illustrates how chosen filters are displayed within the interface, providing a transparent mapping between user selections and the resulting generated images. This design directly responds to participants’ desire for intuitive, visually grounded guidance and aims to reduce the extensive trial-and-error processes characteristic of free-form prompt input.

Figure 2

Filter selection interface and image results page. Multiple style, composition, and mood filters can be combined to refine user prompts visually

In response to the language barriers and conceptual challenges identified in the formative study, the proposed method incorporates a GPT-assisted recommendation system at the Subject stage. Users may input a brief description of the desired image, upon which the system provides five optimized prompt suggestions. These recommendations function to refine, expand, or rearticulate the user’s initial concept, thereby narrowing the gap between vague creative intentions and actionable textual prompts.

The system supports multiple languages, enabling non-native English speakers to convey nuanced ideas without loss of meaning during translation. Figure 3 illustrates the interface through which suggestions are presented, allowing users to accept, reject, or request additional alternatives. By pairing AI-generated textual guidance with explicit user agency, this feature alleviates the steep learning curve associated with traditional prompt formulation while ensuring that core artistic intent remains intact.

Figure 3

GPT-assisted prompt recommendation. The system suggests optimized prompts to address language barriers and refine user ideas

Collectively, these design decisions directly address the needs identified in the formative study, in which participants expressed a strong desire for structured guidance, visual aids, and linguistic support. Consequently, the interface of the proposed method is designed to balance creative freedom with user-friendly assistance, thereby facilitating a more seamless exploration of digital art creation for novice users.

Sixteen participants (7 males, 9 females) with one to five prior interactions with generative AI services were recruited to ensure they remained within the novice range. Users with no prior experience were excluded to avoid confounding factors related to initial system exploration, and frequent users were omitted to preserve the study’s focus on beginner-level challenges. The final sample comprised 10 participants in their 20s and 6 in their 30s, representing a balanced distribution of early-career professionals and graduate students. Table 6 provides demographic information and details regarding prior experience with AI image generators.

Table 6

Demographic analysis of participants in the comparative evaluation

An observational research design was employed in which participants were asked to generate images that closely resembled a reference provided by the researchers. Each participant used two tools: Midjourney and the proposed method. Midjourney served as the baseline system due to its widespread adoption and established user community, making it an appropriate point of comparison. To mitigate order effects, half of the participants began with Midjourney, while the remaining half started with the proposed method. After completing the tasks, participants completed the Computer System Usability Questionnaire (CSUQ) for both services. In-depth interviews were subsequently conducted to collect qualitative feedback regarding overall user satisfaction, perceived creative support, and suggestions for future refinement. The study was conducted between December 1 and December 31, 2023, and each session lasted approximately 70 minutes.

The primary task required participants to generate images that reflect key visual attributes of a shared reference. For Midjourney, participants manually composed text-based prompts, whereas for the proposed method, they followed the step-by-step interface. Observational data were recorded, including strategies used to refine prompts, the time taken to achieve a satisfactory result, and any obstacles encountered. Figure 4 presents the reference image that participants attempted to replicate.

Figure 4

Image presented by the research team

This mixed-methods approach enabled the collection of both quantitative metrics (e.g., usability scores) and qualitative insights (e.g., creative workflow, perceived control), providing a comprehensive understanding of the proposed method’s influence on creative efficacy.

Participants were instructed to generate an image depicting a boy viewed from behind as he observed the stars, using both Midjourney and the proposed method. This task revealed clear differences in how each platform supported creative control and facilitate alignment between intended visual attributes and generated outputs. Table 7 presents outputs from three representative participants, illustrating how the structured, filter-based workflow of the proposed method frequently produced results that more closely aligned with the intended visual concept.

Table 7

Outputs created by participants in the task

For Participant A, Midjourney generated an image that did not reflect the intended sitting posture, instead depicting a standing figure. In contrast, selecting the “Back View” and “Sitting” options within the Composition category of the proposed method resulted in an output that accurately captured the target posture.

Participant B’s output in Midjourney also diverged from the intended atmosphere; the prompt “star fall at night” was misinterpreted as a seasonal reference, producing an autumn-themed scene rather than one characterized by vivid blue night tones. By applying the Makoto Shinkai–inspired filter under the Artistic Reference category in the proposed method, Participant B achieved the desired stylistic and atmospheric qualities with minimal refinement.

A similar pattern was observed for Participant C. The Midjourney prompt “man sitting on the huge tree” yielded an image showing a standing silhouette beside the tree, reflecting ambiguity in prompt interpretation. In contrast, the proposed method’s step-by-step selection process guided the participant toward choices that produced an output closely matching the reference image in terms of shape, composition, and spatial arrangement.

Across all participants, the proposed method’s filter-based interface consistently produced more accurate and detailed images than Midjourney’s open-ended prompt approach. The step-by-step selection of key attributes appeared to bridge the gap between abstract ideas and concrete visual outputs, offering novices a clearer pathway toward achieving their creative goals.

To quantitatively evaluate how effectively Midjourney and the proposed method support novice users, the Computer System Usability Questionnaire (CSUQ) (Lewis, 1995) was administered. Sixteen of the original 19 items were selected to align with the characteristics of web-based services, while three items related to error recovery and help documentation were excluded because such features were not implemented in the current prototype of the proposed method. Participants rated each of the 16 items on a 7-point Likert scale, providing numerical assessments of system usability, information quality, and interface design.

Table 8 presents the average scores for both systems. Across all items, the proposed method consistently outperformed Midjourney, with the most pronounced difference observed for Q12 ("The organization of information on the system screens is clear"), where Midjourney received a mean score of 3.88 compared to 6.13 for the proposed method. These findings correspond with the observational analysis, which indicated that participants experienced difficulty navigating Midjourney’s interface—particularly the reliance on the "/" character and command-line workflow—whereas the structured, step-by-step filter selection supported more intuitive interaction and learning within the proposed method.

Table 8

CSUQ evaluation results for Midjourney and the Proposed Method

To examine whether the observed differences in CSUQ scores between Midjourney and the proposed method were statistically significant, a paired-samples t-test was conducted for each questionnaire item. The analysis revealed that the proposed method achieved significantly higher usability ratings than Midjourney across the majority of items (p < .05).

These results indicate that the differences reported in Table 8 are not attributable to random variation, but rather reflect systematic improvements in usability, information clarity, and interface organization afforded by the proposed method.

Figure 5 presents a comparative analysis across four key dimensions—system usefulness, information quality, interface quality, and overall satisfaction. The proposed method achieved higher scores in all dimensions, underscoring its user-friendly design and effective organization of information. Notably, Midjourney’s overall satisfaction score (5.06) was higher than might be expected given its interface-related challenges, suggesting that users continued to value the quality of its generated images. In contrast, the proposed method’s comparatively lower interface quality score (5.90) indicates areas for refinement, such as enhancing on-screen navigational cues. Overall, these results validate the effectiveness of the proposed method’s guided prompt approach while highlighting opportunities for further improvements to better support novice users.

Figure 5

CSUQ survey score comparison for Midjourney and the proposed method

Following the usability evaluation, participants took part in in-depth interviews to elaborate on their experiences with Midjourney and the proposed method. A thematic analysis (Braun & Clarke, 2006) of the interview transcripts was conducted to examine how each platform supported or constrained users’ creative processes. Four themes emerged, reflecting perceptions of the proposed method’s structured approach, its support for detailed prompt creation, areas requiring interface refinement, and broader implications for digital creativity.

① Effectiveness of step-by-step format and predictable outcomes

Participants consistently emphasized that the step-by-step interface helped translate abstract ideas into more concrete prompts. Novice users who lacked familiarity with technical command structures found this guided approach particularly beneficial. As P2 noted, “The simple interface of the proposed method allowed me to quickly learn its features, and by selecting the desired effects step-by-step, I was able to create images easily.”

Example images accompanying each filter further enhanced predictability and reduced trial-and-error. P12 explained, “By looking at the filter images, I could better predict how my final output would look.” These insights underscore the importance of clear visual cues in reducing uncertainty and fostering user confidence during creative exploration.

② Support for detailed and professional prompt creation

Many participants reported difficulty articulating stylistic or compositional details when using Midjourney’s open-ended text prompts. The proposed method’s category-based structure offered a more accessible way to incorporate complex artistic elements such as Composition, Tone, and Mood. P1 stated, “Midjourney made it hard to think of the right words for style and tone, but with the proposed method, I could easily choose professional terms that refined my prompt.”

The GPT-assisted recommendation system further supported users in expanding vague or partially formed ideas, allowing them to establish clearer mental representations before generating final outputs. As P10 described, “The AI’s suggestions helped me shape my thoughts more clearly, so I could focus on the creative aspects rather than worrying about phrasing.” These findings demonstrate how structured guidance can empower novices to express nuanced creative intentions without requiring extensive expertise.

③ Need for improved interface intuition

Several participants identified opportunities to enhance the intuitiveness of the interface. P3 reported difficulty locating the arrow button for advancing to the next filter, pointing to a need for clearer navigation cues. P14 similarly noted that additional guidance for the Subject input field would help users determine the appropriate level of descriptive detail.

Ambiguities in filter names also posed challenges. P13 remarked, “I wasn’t sure how ‘Cinematic Light’ differed from ‘Dramatic Light,’ and it would help to have more descriptive visuals.” These observations highlight the need for improved clarity in filter labeling and more immediately recognizable interface elements to support seamless interaction.

④ Implications for digital creativity

Taken together, these themes highlight the capacity of structured, visually guided systems to support novice users in expressing complex artistic ideas. By reducing the guesswork associated with prompt formulation, the proposed method enabled participants to focus more on creative decision-making rather than on technical phrasing or command syntax.

These qualitative findings are consistent with prior research that has identified key limitations of free-form text prompting for novice users. Evirgen et al. (2024) report that iterative trial-and-error in prompt refinement imposes a substantial cognitive burden, while Morris (2024) argues that unconstrained text prompts often obscure system affordances and hinder users’ understanding of how inputs translate into outputs. The difficulties observed when participants used Midjourney—such as ambiguity in expressing composition, posture, or stylistic intent—closely align with these earlier findings.

In contrast, the effectiveness of the proposed method’s step-by-step and visually guided interaction reflects insights from prior work on structured guidance and human–AI collaboration. Research on interactive creative systems suggests that decomposing complex tasks into explicit stages can reduce cognitive load and improve user confidence, particularly for beginners (Grossman et al., 2009; Amershi et al., 2019). The present study extends this body of research by demonstrating that category-based filters and visual exemplars can further enhance predictability and perceived creative control in AI-driven image generation contexts.

Unlike earlier prompt-support approaches that primarily emphasize linguistic optimization or command refinement, the proposed method foregrounds visual and conceptual mediation between user intent and system output. This interface-centered approach suggests that visual guidance can complement language-based prompt engineering, particularly during early-stage creative exploration. Overall, these findings indicate that future digital creativity tools should balance intuitive, guided interactions with opportunities for deeper artistic exploration, enabling novice users to pursue their creative visions with greater clarity and confidence.