Creative professionals rarely design for themselves--they design for audiences whose preferences they must anticipate. Yet current text-to-image exploration tools derive diversity entirely from the designer's own input--their prompts, their chosen dimensions, their search queries--confining exploration to what the designer already knows to look for. We present FocusGen, an interactive system that introduces external perspectives into visual design exploration through a "virtual focus group" of simulated persona agents. In contrast to prior persona systems in which multiple agents converge as critics on a single evolving artifact, FocusGen uses personas as parallel generators: each agent--constructed from demographic data, a procedurally generated backstory, and aesthetic preferences elicited through interviews--independently drives an iterative generation loop that produces its own visual concept, transforming one design brief into a spectrum of audience-conditioned directions. With real human participants, we confirm that the iterative refinement loop produces outputs people prefer over zero-shot generation. With synthetic agents at scale, we show that persona conditioning yields higher visual diversity than a generic-assistant baseline--measured by CLIP distance and corroborated by human perceptual judgments--and that open-ended preference interviews yield more diverse outputs than structured ones for both human and synthetic cohorts, while also revealing that agent cohorts recover only part of the diversity of comparable human cohorts. A qualitative study with 16 creative professionals suggests FocusGen helps designers discover unanticipated directions, overcome fixation, and probe audience contexts--while surfacing stereotyping risks that we analyze. We position FocusGen as a divergence scaffold for early-stage ideation rather than a substitute for audience research.
Figures & tables
Figure 1. Comparison of Generative Exploration Paradigms. (Top) Current Approaches: (A) Prompt Optimization requires training, which is computationally intensive and difficult to scale. (B) Dimensional Reasoning relies on user-defined axes (e.g., Lighting, Ambiance), confining exploration to what the designer already knows to look for. (Bottom) Our Approach: (C) FocusGen introduces external perspectives through persona agents equipped with demographics, backstories, and aesthetic preferences. By conducting open-ended preference interviews , the system elicits dimensions the designer may not have considered. Each agent independently generates visual concepts conditioned on its distinct profile, surfacing diverse directions—the designer’s blind spots—to support creative professionals in divergent exploration.
Figure 2. The Feedback Exploration View . (1) The navigation panel lists all past experiments for easy access. When an experiment is selected, its (2) agent configuration and (3) design brief are displayed. (4) The main gallery populates with a mood board of images from the persona crowd. (5) Clicking an image opens a detail view, revealing (6) the iterative refinement history, (7) the agent’s full textual feedback and interview responses, and its demographic profile. A screenshot of the 'Feedback Exploration View,' annotated with seven numbered labels. Label 1 points to the 'Past Experiments' navigation panel. Labels 2 and 3 indicate the summary panels for the agent configuration and design brief. Label 4 shows the main gallery of final generated images. An arrow, labeled 5, shows that clicking an image opens a detail view. Within this view, label 6 points to the iterative history as a series of thumbnails, and label 7 points to the detailed textual information, including the agent's interview responses.
Goal: Design the ideal bedroom.
Format A: Structured Questions
1
Ideal colors and materials for walls, floor, ceiling.
2
Ideal bed (shape, frame, bedding colors, textures).
3
Other furniture or items to include.
4
Type of lighting (ceiling, bedside, accent).
5
Window coverings (curtains, blinds, material).
Table 1. Example interview questions for a bedroom design task. Designers craft their own questions, which can range from specific (Format A) to open-ended (Format B). This choice significantly affects output diversity (Evaluation 2). A compact table showing two interview formats for a bedroom design task. Format A lists five specific questions about colors, bed, furniture, lighting, and windows. Format B uses a single open-ended prompt asking for five preferences in order of importance.
Humans
Agents
Task
Struct.
Open
Struct.
Open
UI Design
.27 (.10)
.34 ( .14 )
.24 (.10)
.29 ( .12 )
Packaging
.20 (.09)
.26 ( .11 )
.21 (.09)
.24 ( .10 )
Logo
.20 (.09)
.25 ( .11 )
.09 (.03)
.12 ( .05 )
Interior
.20 (.08)
.25 ( .10 )
.10 (.04)
.14 ( .06 )
Food
.13 (.05)
.21 ( .10 )
.09 (.03)
.14 ( .06 )
Table 2. Open-ended interviews produce higher visual diversity for both humans and synthetic agents. Values show CLIP Distance with Dispersion in parentheses. A table comparing CLIP Distance and Dispersion for Structured versus Open-Ended interviews across five tasks for Human and Agent cohorts.
Task
Generic
Persona
Δ (%)
UI Design
.22 (.11)
.29 ( .12 )
+32% (+9%)
Packaging
.18 (.09)
.24 ( .10 )
+33% (+11%)
Logo
.08 (.03)
.12 ( .05 )
+50% (+67%)
Interior
.05 (.02)
.14 ( .06 )
+180% (+200%)
Food
.07 (.03)
.14 ( .06 )
+100% (+100%)
Avg
.12 (.06)
.19 ( .08 )
+58% ( +33% )
Table 3. Persona agents produce higher visual diversity than generic baselines across all tasks. Values show CLIP Distance with Dispersion in parentheses. Δ is computed from the two-decimal values shown. A table comparing CLIP Distance and Dispersion for Generic versus Persona agents across five tasks.
Figure 3. Overall usability survey questionnaires and results for our system FocusGen . (Q7 and Q11 are exploratory items; see Section 5.3.) A chart visualizing the results of a 7-point Likert scale survey administered to 16 participants regarding their experience with FocusGen. The results show a high mean score for novelty (Q7, Mean=6.1) and usefulness when creatively stuck (Q10, Mean=5.7), moderately positive scores of approximately 4.5 to 5.6 for most other items, and a near-neutral score for the exploratory comparative item about prompting (Q11, Mean=3.9).
Figure 4. Comparison of concepts generated for the “Coffee Brew 2025” poster task. Concepts from baseline tools (left) were often conventional, while FocusGen (right) generated more diverse and contextually-aware ideas, such as the caffeine molecule and the coffee drop on the city landmark, which participants found novel. A two-panel figure comparing poster designs for a 'Coffee Brew 2025' coffee festival. The left panel, from a baseline tool, displays 12 posters in a similar, conventional illustrative style featuring coffee cups. The right panel, from FocusGen, displays 12 posters in a wide variety of styles, including photorealism and abstract graphics. The FocusGen concepts are more diverse, featuring novel ideas like a city landmark made of coffee and a representation of a caffeine molecule.
Figure 5. An example of how participant P1 used FocusGen to build creative confidence: P1’s initial concepts, generated with a baseline tool as shown in 5(a) , were validated when FocusGen ’s agents independently produced similar ideas as shown in 5(b) . This confirmation gave participant P1 the confidence to then use the system to explore more adventurous, novel concepts (in 5(c) ). A three-panel figure illustrating a participant's creative workflow. Subfigure (a) shows the participant's initial concepts generated using participant's preferred AI tool: models in simple, solid-colored jackets. Subfigure (b) shows a grid of stylistically similar jackets generated by FocusGen, which validated the idea. Subfigure (c) shows a subsequent grid of jackets from FocusGen that are much more adventurous, featuring bright, multi-colored patterns. The figure demonstrates how initial validation can lead to more creative exploration.
Input Prompt Type
Prompt Content
Token Count
Demographic Profile
• Name: Mia • Age: 29 • Gender: Female • Race: White alone • Education: graduate degree • … • Place of Birth: New Mexico/NM
≈ 150
Backstory
I was born 29 years ago in New Mexico, a state that always felt vast and quiet, painted with colors of red earth and infinite sky. My childhood wasn’t particularly eventful, at least not in the dramatic sense. It was… Growing up in New Mexico was a calm existence. We… My focus was always on my studies. High school was a period of intense academic application. I thrived on… When it came time for college, I moved away from New Mexico, drawn by a program that offered a rigorous academic curriculum. I pursued a Bachelor’s degree, followed quickly by a Master’s degree in business, with a specialization in risk management and finance… It was during my Master’s program that I met my husband, Alex… After graduating, I went straight into the insurance industry… I spend my days sifting through information, scrutinizing facts, determining coverage and liability…
≈ 1000
Preference Profile
Here are your interview responses for an ideal bedroom: • I would use matte white paint for the walls with one accent wall in muted sage green. The floor would… • I want a low-profile platform bed with a walnut wood frame. The bedding would be… • I’d include a small desk in the corner near the window, a full-length mirror…
≈ 400
Instruction Prompt
Your task is to roleplay as this character and answer questions based on your profile. Your responses should reflect your character’s personality, background, and experiences.
≈ 200
Table 4. The prompt for a user-simulating persona agent is constructed by concatenating the user’s demographic profile and personal backstory, their elicited preference interview data, and a final instruction for the agent to adopt this profile. A three-column table showing the structure of a prompt for a user-simulating persona agent. The columns are 'Input Prompt Type,' 'Prompt Content,' and 'Token Count.' The table has four rows detailing the prompt's components: a 'Demographic Profile' with example demographic profile (e.g., Name, Age, Gender, Race, Education, Place of Birth) and around 150 expected token count, a 'Backstory' with an example backstory of the agent (e.g., I was born 29 years ago in New Mexico, a state that always felt vast and quiet, painted with colors of red earth and infinite sky. My childhood wasn't…) and about 1000 expected token count, 'Task Preference Data' with example aesthetic preferences with around 400 expected token count, and a final 'Instruction Prompt' commanding the AI to roleplay with expected token count of 200.
ID (Gender, Age)
Profession/Major
YoE
Industry/Major (Employment)
Preferred Gen AI Tools*
P1 (M, 30s)
Brand Designer
8
Finance (Full-time)
Adobe AI , Shutterstock AI
P2 (F, 20s)
Art Director
3
Ad Agency H (Full-time)
ChatGPT , Midjourney
P3 (F, 30s)
Art Director
10
Ad Agency H (Full-time)
ChatGPT
P4 (F, 30s)
Art Director
10
Ad Agency D (Full-time)
ChatGPT , Midjourney
P5 (F, 30s)
Brand Designer
11
Ad Agency C (Full-time)
ChatGPT , Adobe AI, Midjourney
P6 (F, 30s)
Art Director
10
Ad Agency H (Full-time)
ChatGPT , Midjourney, Adobe AI
Table 5. Summary of the 16 creative professionals who participated in our user study. The table details their demographic and professional backgrounds, years of experience, and preferred generative AI tools. The tool indicated in bold is the one each participant chose to use for the baseline condition in our study. A table summarizing demographic and professional information for the 16 participants in the user study. The table has five columns: ID (including gender and age), Profession/Major, Years of Experience, Industry/Major, and Preferred Gen AI Tools. The participants are primarily creative professionals such as Art Directors and Brand Designers, with experience ranging from 3 to 11 years. The most frequently cited AI tool is ChatGPT. The caption clarifies that the bolded tool in the list is the one each participant used for the study's baseline condition.
No.
Task Name
Task Overview
1
Logo Design Ideation
Collect logo references for a new fintech service, “Future Pay” . The core values are “Transparency, Speed, and Trust” .
2
Character Design Ideation
Collect references for the mascot of the “Beauty Expo 2025” . The character will be used to expand the brand’s tone on packaging, app icons, and social media.
3
Clothing/Shoe Design Ideation
Collect outerwear references for the “Urban Run 2025” sports festival’s capsule collection. The main item is a unisex hybrid jacket for both urban commuting and outdoor activities.
4
Ad Poster Design Ideation
Collect poster references for the “Coffee Brew 2025” coffee festival. The poster should emphasize “simplicity + a stylish lifestyle” and will be used in cafes, at bus stops, and on social media.
5
Social Media Campaign Ideation
Collect visual references for the “Summer Spark 2025” social media campaign. Ads will run on Instagram and TikTok, needing a mood that evokes “coolness, trendiness, and a desire to share” .
6
Freestyle Design Ideation
Assume you receive an “open brief” from a new client entering a new market. The client wants to visually explore their brand’s tone and mood.
Table 6. An overview of the six design ideation tasks used in our user study. The tasks were designed to cover a diverse range of creative scenarios, each with a unique brief and set of constraints. A three-column table outlining the six design ideation tasks used in the study. The columns are 'No.,' 'Task Name,' and 'Task Overview.' The six tasks are: 1. Logo Design for a fintech service, 2. Character Design for a beauty expo, 3. Clothing Design for a sports festival jacket, 4. Ad Poster Design for a coffee festival, 5. Social Media Campaign visuals, and 6. a Freestyle Design task with an open brief.
Figure 6. A salient example of demographic filters influencing conceptual direction from the “Beauty Expo 2025” mascot task. Left: Images from agents sampled from Gender A only. Right: Images from agents sampled from all genders. A two-panel figure showing character designs for a 'Beauty Expo 2025' mascot from two participants, P9 and P12. The left panel, labeled 'Gender A,' displays six cute, organic-style characters, including aliens and a sentient mushroom. The right panel, labeled 'Gender B,' displays six characters that incorporate a robotic or technological theme, such as a robot bee and an alien with multiple robotic arms. The figure illustrates how changing the gender filter introduced the concept of robotics into the designs.
Figure 7. The interface for the comparative study in our diversity evaluation. Participants were shown two 3x3 grids of images for a given creative task and were asked to select which grid exhibited greater visual diversity, or to choose a tie. For each pair, one grid was generated using the Persona-Disabled (Baseline) condition, while the other was generated using the full, Persona-Enabled FocusGen system. A screenshot of the user interface for the diversity evaluation study. The interface displays the question, "Which set of images do you think is more diverse?". Below the question are two 3x3 grids of images labeled 'Set A' and 'Set B', both showing various photos of sushi. The images in Set A are visually diverse in composition and type, while the images in Set B are more visually similar to one another. At the bottom, three radio buttons allow the user to choose 'Set A', 'Tie', or 'Set B'.
Vibe design agents translate natural-language requests into interfaces, yet repeated generations often converge on similar visual directions. We present an inference architecture that makes design exploration an explicit stage before implementation. A pre-pass proposes structured design specifications with typicality weights, an external policy samples a direction, and the generator implements it under fixed decoding settings. Candidate elicitation draws on Verbalized Sampling. Applied to UI themes and visual-asset prompts, this architecture provides direct control over aesthetic exploration without changing code-generation temperature. Across 168 benchmark prompts, theme sampling expands observed selection coverage from one option to nearly three and increases rendered visual variation. For visual-asset prompts, the strongest observed LLM-judge win/loss ratio is 1.32, with only a small change in whole-screen similarity. Localized aesthetic choices can thus influence judged preference even when the overall screen remains similar. We further evaluate joint theme and asset sampling in a randomized online study conducted over nearly two months, covering 3.39 million tasks and 87,416 labeled conversations. These studies show how explicit design specifications support exploration of alternative UI concepts and help us evaluate their effects on rendered interfaces, judged preference, and user interaction.
Yifan Zhang, Nghi D. Q. Bui, Georgios Evangelopoulos +1
Professional designers work from client briefs that specify goals and constraints but often lack concrete design details. Translating these abstract requirements into visual designs poses a central challenge, yet existing tools address specific aspects or induce fixation through complete outputs. Through interviews with six professional designers, we identified how designers address this challenge: first structuring ambiguous requirements, then exploring individual elements, and finally recombining alternatives. We developed Brief2Design, supporting this workflow through requirement extraction and recommendation, element-level exploration for objects, backgrounds, text, typography, and composition, and flexible recombination of selected elements. A within-subjects study with twelve designers compared Brief2Design against a conversational baseline. The structured approach increased prompt diversity and received high ratings for requirement extraction and recommendation, but required longer generation time and achieved comparable image diversity. These findings reveal that structured workflows benefit requirement clarification at the cost of efficiency, informing design trade-offs for AI-assisted graphic design tools.
Language models often produce homogeneous responses to open-ended tasks; such homogeneity can spawn groupthink-the convergence of ideas toward a singular and potentially suboptimal decision. We formulate persona diversification as a set-level conditioning problem and study two orthogonal design choices: selecting versus generating personas, and space-filling versus frontier-seeking diversity. We instantiate this design space with four methods spanning coverage and dispersion subset selections, uniform-coverage sampling, and evolutionary persona generation. Evaluations on the Alternative Uses Task (AUT), Infinity-Chat, and Divergent Association Task (DAT) show the benefits of the proposed methods across tasks and creativity objectives. On AUT, evolutionary persona generation increases response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and holistic creativity by 13.9% over task-only prompting, while maintaining 98.5% validity; on Infinity-Chat, it nearly doubles persona-induced response separation relative to random personas. Moreover, evolutionary personas compose with creativity-optimized prompting, further increasing its response diversity by 18.6% and creativity by 6.3%. These results establish persona-set geometry as a task-agnostic mechanism for eliciting divergent LLM outputs, and support persona diversification as a reusable complement to prompt optimization.
Sang Bin Moon, Nicole Cho, Daniel Borrajo +2
School of Electrical and Computer Engineering Purdue University West Lafayette, IN 47909, USA · J.P. Morgan AI Research New York, NY 10172, USA