PaperlayerView exact revision on arXiv

Source-linked AI summary

AIdeation: Designing a Human-AI Collaborative Ideation System for Concept Designers

Wen-Fan Wang, Chien-Ting Lu, Nil Ponsa Campanyà, Bing-Yu Chen, Mike Y. Chen

arXiv:2502.14747v1cs.HC

TL;DR

Concept designers need AI support that fits research-heavy, iterative early ideation rather than only one-step image generation. The paper studies these workflows, develops AIdeation, and reports improved creativity, efficiency, and satisfaction, while identifying limits in detailed control and evaluation rigor.

  • Problem

    Concept designers’ early ideation requires research, brainstorming, and visual exploration, but existing AI tools are not adapted to this workflow and can require complex prompts or produce inaccurate references.

  • Method

    The authors conducted formative research with 12 designers, developed AIdeation for brainstorming, flexible reference search and recombination, and evaluated it through a 16-designer study and a four-studio field deployment.

  • Results

    AIdeation significantly improved reported creativity, ideation efficiency, satisfaction, and enjoyment versus the original workflow, while maintaining comparable idea quality.

  • Takeaways & Limitations

    AIdeation supported broad and deep idea exploration in professional workflows, with two studios continuing to use it for commercial projects after the field study.

  • Takeaways & Limitations

    AIdeation provides high-level rather than detailed control over lighting, atmosphere, camera angles, and composition, and the summative evaluation relied mainly on self-reported data.

Abstract

from arXiv · show

Concept designers in the entertainment industry create highly detailed, often imaginary environments for movies, games, and TV shows. Their early ideation phase requires intensive research, brainstorming, visual exploration, and combination of various design elements to form cohesive designs. However, existing AI tools focus on image generation from user specifications, lacking support for the unique needs and complexity of concept designers' workflows. Through a formative study with 12 professional designers, we captured their workflows and identified key requirements for AI-assisted ideation tools. Leveraging these insights, we developed AIdeation to support early ideation by brainstorming design concepts with flexible searching and recombination of reference images. A user study with 16 professional designers showed that AIdeation significantly enhanced creativity, ideation efficiency, and satisfaction (all p<.01) compared to current tools and workflows. A field study with 4 studios for 1 week provided insights into AIdeation's benefits and limitations in real-world projects. After the completion of the field study, two studios, covering films, television, and games, have continued to use AIdeation in their commercial projects to date, further validating AIdeation's improvement in ideation quality and efficiency.

1 INTRODUCTION

Concept designers need intensive, iterative support for early ideation, while existing AI tools inadequately support their research, exploration, and control needs. AIdeation addresses these needs through human-centered brainstorming, reference search, recombination, and refinement, improving reported creativity, efficiency, satisfaction, and enjoyment.

  • Motivation: Early ideation requires research, brainstorming, and exploration, but designers face difficulty finding aligned references, tight deadlines, and limited breadth and depth.These constraints make it difficult to produce multiple unique designs incorporating diverse elements.
  • Motivation: Existing GenAI tools often require difficult prompt engineering, provide inaccurate or insufficiently detailed outputs, and offer limited iterative control.These issues hinder designers’ ability to use AI during exploratory stages and refine specific visual elements.
  • AIdeation: AIdeation supports breadth exploration by generating diverse visual design ideas from natural language, images, or both.The system presents variations visually so designers can quickly compare directions and continue exploration.
  • AIdeation: AIdeation supports research and depth exploration by extracting design elements as keywords linked to corresponding search results.This connects generated ideas with reference-gathering activities.
  • AIdeation: AIdeation enables iterative refinement through additional references and natural-language instructions while maintaining design diversity.Designers can make broad or focused adjustments aligned with their creative intent.
  • Evaluation: p = 0.001 for creativity preference, p = 0.003 for ideation efficiency, and p = 0.005 for both satisfaction and enjoyment versus the original workflow.The within-subjects study included 16 professional environment concept designers; idea quality remained comparable.
  • Evaluation: Four studios used AIdeation in commercial projects for one week; all reported improved creativity, three reported improved efficiency and quality, and two continued using it afterward.The field deployment involved eight professional environment concept designers.

2 RELATED WORK

Related work frames design ideation as an iterative process and explores tools for visual search, reference recombination, and human-AI collaboration. The paper identifies a remaining gap: no existing tool fully supports concept designers’ complex workflow from research through refinement.

  • Ideation processes: Design ideation alternates between divergent exploration and convergent refinement, using visual searches, reference boards, sketches, and stakeholder evaluation.This process requires repeated movement between generating possibilities and selecting and refining promising directions.
  • Ideation processes: Prior systems support iterative ideation through frameworks such as Wizard of Oz and Muse, while AIdeation combines divergent and convergent support for concept design.The stated aim is to preserve creativity and exploration while integrating current GenAI tools.
  • GenAI design tools: General image-generation tools transform text prompts into visuals but are not specifically adapted to designers’ creative processes.Related work therefore examines improved user experience and prompt refinement.
  • GenAI design tools: GenQuery supports expressive visual search and iterative image-query refinement, while CreativeConnect assists reference recombination through automated suggestions.These systems address distinct research and combination needs within visual ideation.
  • Human-AI collaboration: Human-AI creativity research emphasizes designer curation, iterative communication, nonlinear exploration, and AI roles beyond generation and evaluation.Optimuse and COFI illustrate frameworks for flexible collaboration and balanced divergent-convergent processes.
  • Human-centered AI: Human-centered AI systems aim to augment expertise while providing transparency, explainability, customization, and user control.Examples across creative, medical, and aviation domains align AI contributions with established workflows.
  • Research gap: No existing tool fully addresses concept designers’ specialized workflow across research, brainstorming, and iterative refinement.This gap motivates a system tailored to the complexities of concept design.

3 BACKGROUND: WORKFLOW OF THE ENTERTAINMENT INDUSTRY AND CONCEPT DESIGNERS

Entertainment production moves from development through pre-production, production, and post-production, with concept designers shaping visual direction especially during pre-production. Their workflow separates early ideation from final concept refinement.

  • Industry workflow: Entertainment production comprises development, pre-production, production, and post-production, with concept designers pivotal across the first three stages.Their work is particularly concentrated in pre-production.
  • Concept design workflow: The early ideation phase involves brainstorming and exploring initial ideas through research, visual searches, reference collection, and preliminary sketches.Designers iterate until art directors or clients approve a direction.
  • Concept design workflow: The final concept phase refines approved sketches into detailed designs with depth, texture, and fine details for production teams.These designs guide 3D modeling or set construction and may receive ongoing designer support.
  • Paper focus: This paper focuses on early ideation because it establishes a project’s creative vision, direction, style, and coherence.The stage is described as highly creative and demanding.
  • Concept designers’ role: Concept design spans environments, characters, and props across films, television, and games, translating ideas into assets for downstream production.Examples include Star Wars, DC Comics, Mad Max: Fury Road, and Genshin Impact.
  • Environment workflow: An environment concept designer starts from a client or art-director specification, researches and brainstorms, presents sketches and references for feedback, then refines an approved design for other teams.The workflow is iterative rather than a single generation step.

4 FORMATIVE STUDY

The formative study examined environment concept designers’ workflows, tools, and ideation challenges. Findings informed design goals for AIdeation centered on breadth, depth, grounded information, and controllable iteration.

  • Study design: The formative study investigated workflows, ideation processes, and traditional and AI-based tool use among professional environment concept designers.The broader research program recruited 22 professionals across three studies.
  • Study design: Researchers used 1–2 hour interviews and workflow analyses covering typical projects, research, brainstorming, references, client presentation, and encountered challenges.Transcripts were thematically analyzed using a designer-informed coding framework reviewed by two art directors.
  • Challenges during researching: Designers used search engines, chatbots, Pinterest, portfolio sites, and image databases, but vague or uncommon topics made relevant visual references difficult to find.Traditional search often failed to align with design intentions or provide sufficient material for blending styles and themes.
  • Challenges during brainstorming: Designers typically needed 3–5 environment variations under deadlines of roughly half a day to one day, limiting deeper research and creative development.Complex or unfamiliar settings required substantial effort to generate and combine design variations.
  • Challenges during brainstorming: Although all participants had experience with AI design tools, nine had already integrated them into their workflows.This adoption occurred alongside continuing challenges in research and brainstorming.
  • Problems with current AI design tools: Current AI tools require complex prompts and often demand a clear idea in advance, conflicting with exploratory visual workflows.Participants also reported difficulty obtaining desired outcomes from repeated prompt modifications.
  • Problems with current AI design tools: AI-generated references can contain hallucinations, inaccurate details, unexplained elements, and content that diverges from prompts, reducing their usefulness for grounded design decisions.Participants described many outputs as suitable mainly for mood reference rather than detailed incorporation.
  • Problems with current AI design tools: AI tools also provide insufficient control for changing one element while preserving others, making iterative refinement difficult.These findings motivated design goals for breadth and depth exploration, grounded references, and flexible control.

5 SYSTEM & IMPLEMENTATION

AIdeation supports early concept-design ideation by unifying brainstorming, reference research, and iterative refinement. Its interface and pipeline connect generated ideas, searchable references, and two refinement modes.

  • System goals: AIdeation unifies research, brainstorming, and design-idea refinement into a cohesive, iterative workflow.The system is designed to support breadth exploration, depth exploration through research, and flexible iterative exploration.
  • Brainstorming: The system generates eight design ideas organized across Theme, Contents, Art Style, Lighting and Atmosphere, Color Palette, and Shot Angle.Ideas appear as images with titles in the Ideas Overview Panel and can serve as potential hero references.
  • Research: The Idea Detail Panel extracts keywords into six categories and links selected keywords to related supporting-reference searches.This structure helps designers inspect a concept’s composition while accessing additional information and detailed references.
  • Refining design ideas: Designers can refine an idea by combining it with a selected reference or by giving natural-language instructions.Reference combination generates five variations, while instruction-based refinement preserves the original idea’s essence while introducing diversity.
  • Next ideation cycle: AIdeation supports another brainstorming cycle from the current idea, enabling related tasks such as designing a kitchen based on an existing environment.The interface also records an idea’s origin, including combinations of prior ideas and references.
  • Implementation: The technical pipeline captions image inputs, uses language models to generate or modify descriptions, extracts keywords, and renders images with DALL-E 3.Keyword selections trigger Bing Image Search, while selected references and instructions enter separate refinement pipelines.

6 SUMMATIVE STUDY

The summative study compared AIdeation with each designer’s preferred existing workflow in a within-subject study of 16 professional environment concept designers. It evaluated ideation support, ideation quality and efficiency, and workflow support across stages using tasks, questionnaires, and interviews.

  • Study design: The within-subject study examined AIdeation’s support for ideation, ideation quality and efficiency, and workflow stages.The study involved 16 professional environment concept designers.
  • Study design: Each participant’s preferred existing workflow served as the baseline, including image databases, search engines, and AI design tools.Participants without prior AI-design experience received access to ChatGPT-4 with DALL-E 3 and a brief tutorial.
  • Procedure: The study used counterbalanced 30-minute design tasks under each condition, preceded by practice sessions and followed by questionnaires and interviews.The full session lasted 2 to 2.5 hours, with a 10-minute briefing, breaks, and a post-study interview.
  • Tasks: The final tasks required participants to gather at least three reference sets for interior and exterior designs using PureRef.The design topics included a Mayan Observatory and Planetarium and a Tibetan Meditation Research Center.
  • Task refinement: Sketching was removed after pilot participants reported extreme stress and could not complete the original 40-minute task.Three art directors validated the revised reference-only tasks as common practice under time constraints.
  • Measurements: Measures covered breadth, depth, flexibility, efficiency, idea quality, creativity enhancement, and workflow-stage support.Participants also explained their questionnaire ratings in interviews, while external expert evaluation of collected references was not conducted.
  • Participants: The sample included 16 designers from animation, games, art outsourcing, and freelancing, with 1–12 years of professional experience.Participants came from five studios, and their mean experience was 4.6 years (SD = 3.2).

7 RESULTS & FINDINGS

Compared with participants’ original workflows, AIdeation supported broader, deeper, and more flexible exploration, improved creativity and ideation efficiency, and generally increased satisfaction. Participants also valued its research, reference-gathering, visual presentation, and iterative recombination support, although controllability and exact-match efficiency remained uneven.

  • Breadth, depth, and flexibility: Participants preferred AIdeation for breadth, depth, and flexibility in idea exploration, with 69% preferring it for both breadth and depth.AIdeation supported iterative refinement and reference combination, though some participants reported limited atmospheric or stylistic diversity and detailed control.
  • Creativity: 81% preferred AIdeation for enhancing creativity over their original workflow.Participants attributed this to unexpected brainstorming results, keywords for inspiration, and combinations of diverse or uncommon elements.
  • Overall satisfaction, task efficiency, and difficulty: Task-efficiency findings were mixed because waiting for results and pursuing exact matches sometimes consumed time or prevented participants from completing tasks.Participants who iterated freely and avoided overanalyzing prompts generated more diverse outputs, whereas excessive input refinement reduced efficiency.
  • Research, reference gathering, and visual presentation: AIdeation significantly improved information gathering and visual presentation, with 81% and 69% of participants preferring it, respectively.Keywords, supporting references, and generated images helped participants understand topics, collect relevant information, and communicate ideas to clients or directors.
  • Qualitative findings on AIdeation usage: AIdeation received polarized evaluations of controllability: reference combination and instruction-based refinement helped some users, while others could not preserve layouts or adjust specific properties.Some participants valued lower controllability because it produced significant variation and more elements to extract.

8 FIELD STUDY

A one-week field study across four studios examined AIdeation in ongoing commercial projects, finding reported gains in creativity, quality, and, for some studios, efficiency, alongside styling and controllability limitations.

  • Study Setup: 4 studios used AIdeation in ongoing commercial projects for one week, with diary studies and interviews assessing workflow integration, outcomes, and efficiency.The study involved 8 professional environment concept designers and examined AIdeation alongside existing design tools.
  • Usage: 1,092 ideas across 98 cycles produced 60 selected ideas contributing to 12 environments.The field-study usage is summarized in Table 2 by studio, project, ideation cycles, generated ideas, and selected ideas used in final outputs.
  • Efficiency: Both S3 and S4 reduced design time from 5 days to 2 days and from 14 days to 6 days, respectively, while S2 reported a slight efficiency decrease.S3 and S4 linked the savings to quickly identifying design directions; S2 cited client preference for the generated artistic style and time-consuming image generation.
  • Quality: S1, S3, and S4 reported significantly enhanced final-design quality, citing richer designs and greater diversity than a DALL-E 3 comparison in S3.The reported quality improvements included adding varied design elements to final outputs.
  • Creativity: All four studios reported boosted creativity, including unexpected outcomes and uses such as generating distinctive patterns difficult to source online.One participant also reported obtaining useful ideas from uploading an image without prompts.
  • Challenges: Participants reported styling problems, consistently symmetrical scenes, and limited controllability because changing one element could alter the whole image.They requested gradual generation and inpainting for more detailed control; complex initial designs also required simplification instructions.
  • Continued Usage: Studios 1 and 4 continued using AIdeation in production after the field study, with Studio 1 reporting approximately 40% time savings while creating six scenes in two weeks.Studio 4 created 22 scenes in six weeks, and its participant reported outputs greatly exceeded previous work, although time savings were not measurable under a fixed deadline.

9 DISCUSSION, LIMITATIONS, AND FUTURE WORK

AIdeation addresses adoption barriers by improving transparency, control, and alignment with concept designers’ iterative workflows. Field evidence suggests practical benefits, while limitations remain in study measurement and fine-grained visual control.

  • Addressing barriers: AIdeation makes brainstorming results more transparent by presenting visuals and categorized keywords, helping designers understand and refine creative directions.This reduces reliance on interpreting generated images or manually constructing complex prompts.
  • Field examples: In real-world examples, one designer reached a desired bridge result after two refinements, while another selected five mountain-scene results after 10 ideation cycles but exceeded the original estimate by one hour.The cases illustrate both efficient refinement and a practical time cost in extended exploration.
  • Workflow alignment: Its nonlinear, modular workflow supports iterative ideation and lets designers switch between AI and non-AI tools according to creative needs.The design is intended to align more closely with concept designers’ ideation processes than linear, one-step tools.
  • User perceptions: Designers reported that AIdeation was easier to control and communicate with than other AI tools, particularly for expressing intended modifications.This supports the importance of user control and clear communication for engagement and satisfaction.
  • Limitations: The summative study relied mainly on self-reported data, and the field study lacked quantitative measures and had less experimental control.The authors suggest longer summative sessions focused on narrower tasks as one direction for future research.
  • Limitations: AIdeation provides high-level control but does not yet precisely adjust lighting, atmosphere, camera angles, or composition while preserving other elements.Participants described the system as covering roughly 70–80% of client communication, with detailed control needed for the remaining refinement.
  • Future work: Users also noted limited diversity in art styles, atmospheres, and camera angles because of the image-generation model’s constraints.Future updates could support selectable styles, atmospheres, or fine-tuned models.

10 CONCLUSION

AIdeation combines traditional references with AI-generated outputs to support broad and deep idea exploration. Across professional evaluations, it improved creativity, efficiency, and the diversity of ideas while maintaining comparable quality to original workflows.

  • Conclusion: AIdeation integrates research, brainstorming, reference gathering, and design refinement into one iterative workflow for exploring visual concepts.It combines diverse references with generated outputs based on user input.
  • Conclusion: A study with 16 professional concept designers found significantly broader and deeper exploration, higher creativity, and greater efficiency than original workflows, with comparable idea quality.The conclusion presents these effects across multiple dimensions of ideation rather than as a single task result.
  • Conclusion: A field study in four design studios further indicated potential efficiency gains for complex design tasks.The field evidence extends the evaluation beyond controlled ideation activities into commercial settings.

General Questions

The general-question materials document the interview and questionnaire prompts used to compare AIdeation with designers’ original workflows. They cover efficiency, satisfaction, exploration, creativity, feature usefulness, and future adoption.

  • Acceptance: Participants were asked to rate AI acceptance from completely unacceptable to acceptable for direct use in design.The question distinguishes using AI as a reference or aid from using it directly in design.
  • Workflow comparison: Follow-up questions examined differences from the original workflow, thought processes, reference use, efficiency, and task difficulty.These prompts sought qualitative explanations for participants’ comparative system choices.
  • Feature assessment: Participants were asked which features were most helpful and how AIdeation changed material search, reference gathering, and generation of diverse or high-quality ideas.Additional questions addressed strengths, weaknesses, and the reasons behind comparative ratings.
  • Evaluation dimensions: The questionnaire compared systems on breadth, depth, flexibility, creativity, satisfaction, enjoyment, and task difficulty.These dimensions cover both ideation outcomes and user experience.
  • Future use: The interview also asked whether AIdeation lacked features, needed improvement, or would be used in future work.These questions directly probe perceived limitations and intended adoption.

C Appendix C: Idea Generation GPT

Appendix C specifies an idea-generation assistant that converts instructions, reference descriptions, and creativity scores into structured English design ideas. Its examples organize outputs into consistent visual-design categories.

  • Inputs: The assistant receives an instruction, a reference-image description, or both, together with a creative score between 0 and 1.The creative score indicates how diverse the generated idea should be from the original image.
  • Creativity control: Low creativity scores call for ideas close to the image description, whereas high scores permit more innovative and diverse variations while retaining relevance.The instructions prioritize artist-provided directions regardless of the creativity score.
  • Output format: The output is structured under Theme, Art Style, Content, Lighting and Atmosphere, Color Palette, Layout, and Shot Angle.Content may contain a variable number of subcontent items and the output must be in English.
  • Example outputs: The 1930s film-camera-room example expands a reference into a realistic industrial workspace with equipment, furniture, lighting, palette, layout, and a 3/4 view.Its visual description emphasizes a dark, cluttered setting with focused lighting and vintage details.
  • Example outputs: The fantastical-village example transforms organic architecture, greenery, pathways, characters, and mystical features into a painterly concept-art specification.The example includes a central gourd-like treehouse, connected pathways, natural elements, and magical lighting cues.

D Appendix D: Keyword Extraction GPT

The keyword-extraction component converts design ideas into concise, categorized search terms so designers can retrieve visually relevant references. It emphasizes salient visual elements while constraining keyword counts and format.

  • Keyword extraction: Keyword extraction organizes visual elements into theme, art style, content, lighting and atmosphere, color palette, and shot-angle categories.The instructions require simple, descriptive keywords focused on the design idea’s key elements.
  • Keyword extraction: Keywords are limited by category, including at most 3 for theme, 3 for art style, 20 for content, 5 for lighting and atmosphere, 5 for color palette, and 3 for shot angle.The output also follows a prescribed markdown structure with nested content subcategories.
  • Keyword extraction: The system adds descriptive adjectives, keeps keywords under 5 words, focuses on key elements, and omits the layout category.Examples distinguish terms such as “vintage car” from generic object labels.
  • Example output: The example output groups the extracted terms under theme, art style, content, lighting and atmosphere, color palette, and shot angle.Its content section further divides the room into areas such as the central workstation, darkroom corner, repair station, and furniture.

E Appendix E: Combining Idea GPT

The combining-idea component integrates reference-image descriptions into original design ideas while controlling how much creative variation is introduced. It uses the specified keyword or instruction to guide modifications and preserves the requested design structure.

  • Reference integration: Reference-image descriptions are blended into original design ideas according to a variety score ranging from 0 to 1.A score of 1 signifies a design that significantly diverges from the original, while lower scores favor more limited changes.
  • Reference integration: The keyword identifies which part of the original idea to modify, while the reference description supplies the content of that modification.The keyword should locate the target part without influencing the modification itself.
  • Reference integration: If the keyword is absent, the system finds a reasonable integration point and aims for a harmonious blend between the original idea and reference image.The guidelines require the specified part to be modified regardless of the variety score.
  • Output structure: The output retains structured fields for theme, art style, content, lighting and atmosphere, color palette, layout, and shot angle.Content may contain a variable number of subcontent sections.
  • Example output: The forest-retreat example incorporates stone-covered domes and curved brick pathways into the original lakeside setting while retaining its broader atmosphere and composition.The revised design combines the reference house’s stone construction and landscaping with the original forest, lake, characters, and layered backdrop.
  • Creative variation: A separate creative-score instruction directs the system to make only specified changes at low scores and broader but relevant changes at high scores.Generated outputs must remain below 400 characters.
Loading 2502.14747v1…