cs.HCApr 13, 2026

Brief2Design: A Multi-phased, Compositional Approach to Prompt-based Graphic Design

Authors: Kotaro Kikuchi, Nami Ogawa

Organizations: CyberAgent, Japan

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Figure 1. Brief2Design is an end-to-end system that supports designers in creating graphic designs from design briefs. The system first (a) structures the input text into requirement cards and (b) generates type-specific element cards. After designers (c) select the element cards, the system (d) integrates the selected elements into a single design. The designers can easily explore design variations by recombining element cards, such as changing the composition cards.
Figure 2. Step 1 interface: Structuring free-form briefs into requirement cards. Users input their design brief (a) and trigger requirement extraction (b), which categorizes the input into eight requirement cards (c). Users can then request AI-generated recommendations (d) for each category or manually add custom entries (e).
Figure 3. Step 2 interface: Generating type-specific element cards with preview images. Users initiate element generation (a), and the system displays element cards for five categories: Objects, Backgrounds, Text, Typography, and Composition (b). Cards for visual elements include preview images (c), enabling users to visually evaluate each element. At the category level, users can manually add new candidates (d) or request additional AI recommendations (e). Users can then select their preferred elements (f) for integration in the next step.
Figure 4. Element card interface in Step 2. The left shows an element card with Edit, Delete, and Regenerate buttons. The right shows the editing interaction flow where users can modify the rough prompt (removing “holding neon glow sticks” in this example) and update the preview image.
Figure 5. Step 3 interface: Integrating selected elements into a final design. Users review their selected elements (a) from Step 2 and initiate integration generation (b). The system then displays the generated final design (c), the integrated prompt (d), and the used elements (e). Users can explore variations through the generation history (f).
Figure 6. Baseline system interface: A conversational AI interface similar to ChatGPT. Users interact with the AI through text prompts to generate images. The interface features a chat-based interaction model with message history and image generation capabilities.
Figure 7. Examples of submitted images from different participants (P2, P11, P7, and P6) under different task conditions (low-freedom and high-freedom) and systems (Brief2Design and Baseline).
Figure 8. Image diversity comparison for generated and submitted images (ns: not significant). The circle markers indicate data points, and the cross markers indicate means.
Figure 9. Prompt diversity comparison for generated and submitted images (ns: not significant, ∗\ast : p<.05p<.05 , ∗∗\ast\ast : p<.01p<.01 ). The circle markers indicate data points, and the cross markers indicate means.
Figure 10. Left: Scatter plot showing the relationship between prompt distance and image distance for the submitted images. The dotted green lines indicate the medians of each axis. Right: Two example image pairs corresponding to representative points (a) and (b).
Figure 11. Results of the Creativity Support Index (CSI) questionnaire.
Figure 12. Rating proportions for each feature in the proposed system.
Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

ElementGuideline
ObjectCreate AI image generation prompts for the main subject only (person, product, animal, logo, etc.). Include diverse visual styles (photorealistic, illustration, sketch, etc.). Use concrete visual descriptions, not abstract concepts. Exclude backgrounds and settings.
BackgroundCreate AI image generation prompts for backgrounds only (solid colors, gradients, textures, patterns, scenes, etc.). Include diverse visual styles. Use concrete visual descriptions, not abstract concepts. Exclude foreground objects.
TextProvide actual content using the format “Role: Text content” ( e . g ., “Headline: Summer Sale 50% Off”). Use role labels in the same language as the output language. You may recommend multiple text items for the same role if appropriate.
TypographyDescribe typography for BOTH headings and body text as a complete typography system. Cover type classification appropriate for {output_language} , font personality, visual weight, font sizes (large, medium, small), colors, and special effects. Avoid specific font names.
CompositionDescribe layout approaches using intuitive, visual language with concrete examples like “magazine-style with large hero image at top.” Focus on HOW elements are visually organized. Use abstract terms like “main visual area,” “text zone” instead of describing image content.
Appendix
Table 1. Element-specific guidelines injected into the ElementRecommender prompt via the {element_description} variable.
Table 2. Design outputs collected for T1 (low-freedom). Each row lists two participants from the proposed system (left) and two from the baseline condition (right).
Table 3. Design outputs collected for T2 (low-freedom). Each row lists two participants from the proposed system (left) and two from the baseline condition (right).
Table 4. Design outputs collected for T3 (high-freedom). Each row lists two participants from the proposed system (left) and two from the baseline condition (right).
Table 5. Design outputs collected for T4 (high-freedom). Each row lists two participants from the proposed system (left) and two from the baseline condition (right).
More in this topic

Human-AI Co-Creation

Human-AI co-creation studies how people and AI jointly develop ideas, artifacts, and decisions while preserving meaningful human agency.

See all 153 papers

Explore similar work

CardsList

  1. Graphic-Design-Bench: A Comprehensive Benchmark for Evaluating AI on Graphic Design Tasks

    Apr 5, 2026Adrienne Deganutti, Elad Hirsch, Haonan Zhu +2Vector Graphics Generation

  2. DOGS: Design-Space Sampling for Prompt-Driven Logo Generation

    Oct 7, 2026Ganyu Zou, Chen Dai, Nathan Self +5Image GenerationPrompt Optimization for T2I Generation

    Department of Computer Science Virginia Tech Alexandria, Virginia, USA · Verisign, Inc. Reston, Virginia, USA

  3. TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design

    May 20, 2026Haonan Zhu, Elad Hirsch, Alexandria Minetti +3VLM EvaluationT2I Generation Evaluation

    Lica World, San Francisco, USA. · Contra.Work Inc., Brooklyn, NY, USA.