Exploring the Affordances of Generative Image AI for Supporting Early-stage Architect-client Communication

1 Introduction

• Architectural communication is crucial because stakeholders may interpret the same linguistic symbols differently, risking conflicting goals and misaligned understanding.The introduction links communication failures to project success and clients’ needs.Source: §1 · p. 1
• Earlier communication technologies improved efficiency but often created professional barriers for non-experts, motivating renewed search for supportive tools.The paper notes ICT growth during the 2010s followed by stagnation.Source: §1 · p. 1
• Text-to-image generators can produce high-fidelity visuals from text prompts in seconds, and prior work suggests they may bridge design communication gaps.The paper identifies Dall·E, Midjourney, and Stable Diffusion as examples.Source: §1 · pp. 1–2
• The study examined how generative AI renderings affect ideation and communication in video-conferenced design sessions with 11 architect-client pairs.Pairs used Stable Diffusion XL Beta and then completed semi-structured interviews.Source: §1 · p. 2
• Generative image AI helped pairs build common ground with high-fidelity visuals, while clients could generate images expressing their desires and influence design direction.The preliminary findings draw on early-career architectural students as architect participants.Source: §1 · p. 2

2.1 Challenge of Architectural Design Communication

• In the fuzzy front end, idea generation, selection, and communication compete with time-consuming low-fidelity prototyping; expertise and language gaps also hinder shared understanding.The paper notes that clients may lack technical vocabulary and that architects may rely on their own experience.Source: §2.1 · p. 2

2.2 Practice for Assisting Architecture Design Communication

• Briefs and sketches support requirements and ideation, while BIM and VR can aid coordination but may impose client barriers or require substantial development effort.Sketch ambiguity can encourage innovation while also introducing communication uncertainty.Source: §2.2 · p. 3

2.3 Generative Image AI as Visual Material

• Generative image AI offers concrete, rapid high-fidelity visuals for design conversations, with ambiguity that can support varied interpretations and ideation.The paper also notes that generated 3D meshes remain insufficiently detailed, so some systems use 2D images to inform 3D design.Source: §2.3 · p. 3

3.1 Recruitment Procedure

• Recruitment yielded 11 pairs; architects had at least three years of design education and architecture-related work, while clients were screened to exclude architecture-related work or study.Architect participants were recruited primarily from one anonymized city, while clients came from different U.S. cities.Source: §3.1 · p. 3

3.2 Participants Demographic Information

• Architects averaged 7.18 years of design education and 1.86 years of work experience; client participants averaged 26.27 years old.Most architects were master’s students, and all had at least one year of work experience.Source: §3.2 · p. 4, Table 1 · p. 4, Table 2 · p. 4

4.1 Study Activities

• Pairs met remotely for an 80-minute session, sharing access to image generation and a Figma board while independently creating image variations.The session included a 10-minute introduction, 45-minute design task, and 20-minute interview section.Source: §4.1 · p. 5

4.1.1 Introduction Section.

• Participants reviewed onboarding materials, practiced with DreamStudio, and added a generated sofa image and its prompt to the Figma board.The practice prompt example described a sofa with classical ancient Chinese styles.Source: §4.1.1 · p. 5

4.1.2 Design Section.

• The design task asked clients to request a house by the sea; pairs iterated until satisfied or the 45-minute limit, recording images and prompts on the board.The sea requirement provided a common constraint for comparing sessions; the figure shows the board before and after the design section.Source: §4.1.2 · p. 5, §4.1 · p. 5, Figure 1 · p. 6

4.1.3 Interview Section.

• After the design task, each participant completed an individual semi-structured interview lasting approximately 10 minutes.Source: §4.1.3 · p. 5

4.2 Qualitative Analysis

• Researchers transcribed the interviews, conducted initial and refined coding, and used affinity diagramming to develop themes through discussion.The first author developed initial codes, then two authors independently coded transcripts and adjusted the codes as needed.Source: §4.2 · p. 6

5.1.1 Generative Image AI Provides Solid Visual Materials.

• Generated images gave pairs something concrete to discuss, supported rough shared understanding, and let clients show architects the designs they wanted.Participants also saw images as complements to specification documents containing details such as budget, cost, and room numbers.Source: §5.1.1 · pp. 6–7

5.1.2 Generative Image AI Facilitates More Efficient Communication.

• Participants said image generation made communication more efficient and faster, helping them understand one another and make design decisions more quickly.They described the process as saving communication effort and making it faster to judge what they liked or disliked.Source: §5.1.2 · p. 7

5.1.3 Participants Think AI Sparks Design Inspirations and Provides Unique Designs.

• Participants valued generated images as early-stage inspiration, said they helped clients clarify preferences, and described shared prompts as a way to exchange design vocabulary.Architects used more specific design terms, which clients could adapt; architects also drew information from clients to construct prompts.Source: §5.1.3 · p. 7, §5.1.4 · p. 7

5.2.1 Participants Find It Hard to Generate Images They Want.

• Participants struggled to steer outputs toward their intent, especially for specific requirements such as floor count, facade color, room layout, or budget.They reported ignored prompts and difficulty changing results, while some used instructional rather than descriptive prompts.Source: §5.2.1 · p. 8, §5.2.2 · p. 8

5.2.3 Perceived Bias in Generated Images.

• Participants described outputs as dreamy, fancy, similar-looking, or highly saturated, and said popular styles could constrain exploration and creativity.One participant also reported difficulty generating a requested Chinese-style exterior.Source: §5.2.3 · p. 8

5.2.4 “Architecture is Reduced to 2D Images”.

• Because generations were 2D and could change entirely between rounds, participants worried about feasibility, lost context and agency, and weakened iterative design.Architects and clients could interpret successive images differently, and architects said rapid changes could distract clients and interrupt their train of thought.Source: §5.2.4 · pp. 8–9, §5.2.5 · p. 9

5.3.1 Participants Seek More Precise and Flexible Control over the Outcome.

• Participants sought stronger controls, including hard rules, sketch inputs, and partial generation, alongside conversational image editing and clarification.Suggested conversational systems could ask users questions and offer options when more clarification is needed.Source: §5.3.1 · p. 10, §5.3.2 · p. 10

5.3.3 Facilitating User Personalization and Divergent Thinking.

• Participants proposed learning from user selections, supporting more diverse styles, and integrating generation, sharing, and collaborative prompt modification.They described the existing copy-and-paste workflow between the generator and shared whiteboard as extra work.Source: §5.3.3 · p. 10, §5.3.4 · p. 10

6.1 Generative Image AI Altered Design Communication

• High-resolution, accessible AI renderings give both stakeholders concrete material for discussing preferences and reducing linguistic ambiguity, though whether they inspire rather than fixate design ideas needs further evaluation.The authors distinguish client-preference design from ideation and call for rigorous evaluation of inspiration versus a false perception of it.Source: §6.1 · p. 11

6.2 Generative Image AI Changed Power Dynamics and Beyond

• Clients gained voice and agency through visualizing their goals and sharing prompts, while some architects worried about unpredictability, loss of domain expertise, and design processes reduced to images.The authors caution that building empathy with clients cannot simply be automated and that perceived efficiency gains should be weighed against hidden risks.Source: §6.2 · pp. 11–12

6.3.1 Improving User Control with Symbolic-connectionist System.

• Future systems should improve control by combining interpretable rules with connectionist generation, particularly for architectural specifics such as floor and room counts.Participants wanted rules for the AI to follow; the authors suggest symbolic-connectionist approaches to improve user control.Source: §6.3.1 · p. 12

6.3.2 Adaptive System Design for Different Design Stages.

• Interfaces should adapt to design stage: support preference-driven convergence while encouraging divergent exploration, since image examples can contribute to design fixation.Clients may discover their preferences through image AI, so the authors propose context- and stage-aware interfaces.Source: §6.3.2 · p. 12

6.3.3 Providing Extra Assistance for Prompt Engineering and Collaborative Prompting.

• Collaborative prompting and prompt-engineering aids could streamline shared work and help users refine design intent through templates, prompt lists, auto-revision, or conversation.The authors note that refining prompts is arduous and takes many revisions.Source: §6.3.3 · pp. 12–13

6.3.4 Mitigating the Stylistic Bias in Image Generation.

• Training-data and model biases can favor particular architectural styles, hindering nuanced communication and aesthetic diversity; proposed remedies include fairness measures, bias disclosure, and style fine-tuning.The authors suggest these approaches could better align generated images with users’ communicative intent.Source: §6.3.4 · p. 13

7 Conclusion

• In early-stage design scenarios involving early-career architects and clients, generative image AI supported more efficient communication and shared understanding, while biased image results remained a limitation.The authors describe four directions for future system design for ethically utilizing generative image AI.Source: §7 · p. 13

A.1 Recruitment Questionnaire

• The study invitation sought designers from various domains and people willing to serve as design clients for an 80-minute study.Participants were offered a $15 gift card.Source: §A.1 · p. 16
• The questionnaire asked whether respondents worked in design or wanted to be a client, and requested their design-domain experience.It also asked how many years they had studied or worked in their domain in an English-speaking country.Source: §A.1 · pp. 16–17
• Applicants reported native or bilingual English fluency, or qualifying TOEFL, IELTS, or Duolingo scores, and could provide an email if interested.The listed thresholds were TOEFL 102, IELTS 7.5, or Duolingo 120.Source: §A.1 · p. 17

A.2.1 Text-to-image generation.

• Users enter a prompt in the text box and click Dream to generate images.The accompanying figure depicts DreamStudio’s image-generation interface.Source: §A.2.1 · p. 17, Figure 4 · p. 17
• The DreamStudio interface changed during the study, but its main features, including parameter changes, remained the same.Source: §A.2.1 · p. 17
• Table 3 presents example prompts and corresponding images at different levels of detail.Source: Table 3 · p. 18

• Participants were asked whether they would use the AI tool to facilitate communication between clients and designers.Source: B Interview Questions · p. 18
• The guide asked whether participants expected the tool to improve or worsen design communication and why.Source: B Interview Questions · p. 18
• Participants were asked whether they used one another’s prompts and whether sharing prompts helped their design process.Source: B Interview Questions · p. 18