Source-linked AI summary
Using Text-to-Image Generation for Architectural Design Ideation
Ville Paananen, Jonas Oppenlaender, Aku Visuri
TL;DR
Text-to-image generators remain underexplored as supports for creativity during the early stages of architectural design. The authors study 17 architecture students using three generators and find that image generation can enrich ideation through serendipitous discovery and imaginative engagement when design constraints are carefully considered.
Problem
The paper addresses the limited evidence about how text-to-image generators support divergent creativity during early-stage architectural concept design.
Method
The authors conducted a laboratory study with 17 architecture students designing a culture center using Midjourney, Stable Diffusion, and DALL-E, alongside questionnaires and group interviews.
Results
Image generation could meaningfully support architectural design by enabling serendipitous idea discovery and an imaginative mindset when design constraints are carefully considered.
Takeaways & Limitations
Effective architectural use requires image-generator design that supports creative exploration and education that emphasizes appropriate and advanced usage.
Takeaways & Limitations
The study’s participants were largely inexperienced, so better instructions or extended tasks might improve generated images and resolve observed challenges.
Abstract
from arXiv · showhide
The recent progress of text-to-image generation has been recognized in architectural design. Our study is the first to investigate the potential of text-to-image generators in supporting creativity during the early stages of the architectural design process. We conducted a laboratory study with 17 architecture students, who developed a concept for a culture center using three popular text-to-image generators: Midjourney, Stable Diffusion, and DALL-E. Through standardized questionnaires and group interviews, we found that image generation could be a meaningful part of the design process when design constraints are carefully considered. Generative tools support serendipitous discovery of ideas and an imaginative mindset, enriching the design process. We identified several challenges of image generators and provided considerations for software development and educators to support creativity and emphasize designers' imaginative mindset. By understanding the limitations and potential of text-to-image generators, architects and designers can leverage this technology in their design process and education, facilitating innovation and effective communication of concepts.
1 Introduction
The paper examines how text-to-image generators can support creativity and ideation during the early, exploratory stages of architectural design. It also asks how effective out-of-the-box systems are and what developers should consider for architectural use.
- Generative AI enables detailed architectural concept representations from natural-language prompts, extending traditional visual media used for ideation and communication.
- The study investigates how text-to-image generators can support creativity and ideation during the “fuzzy front end” of architectural concept development.
- It evaluates the effectiveness of out-of-the-box text-to-image generators in architectural design and considers implications for future development.
3. What are the typical challenges of text-to-image generator use and text prompting for novel users?
The paper identifies challenges in using text-to-image generators for architectural ideation, including the need to account for design constraints and users’ imaginative engagement.
- Meaningful image generation requires careful consideration of design constraints and imaginative ideation within the architectural design process.
- Generative tools can support serendipitous idea discovery and an imaginative mindset, enriching architectural design exploration.
2 Related work
Prior work frames architectural creativity as both producing novel, useful outcomes and applying creative skills during design. Although text-to-image systems are intuitive and increasingly capable, their architectural use and real-world benefits remain insufficiently explored.
- 2.1 Creativity in the architectural design process: Architectural creativity is commonly understood through novelty and usefulness, while also involving creative skills applied during the design process.
- 2.1 Creativity in the architectural design process: Text stimuli have been shown to support nonlinear creativity and shift architectural design attention from outcomes toward the creative process.
- 2.2 Text-to-image generation in architectural design: Generative AI research in architecture has grown, but machine learning has more often focused on 3D generation than 2D methods.
- 2.2 Text-to-image generation in architectural design: Text-to-image systems offer a natural-language interface, yet output creativity depends on user skill, prompt keywords, and often longer prompts.
- 2.2 Text-to-image generation in architectural design: Existing studies find broad applicability in built-environment contexts, but architectural semantic ambiguities and real-world benefits remain open questions.
3 Method
The authors conducted a laboratory study in which 17 architecture students designed culture-center concepts with three public text-to-image generators. They combined a structured design task with questionnaires, interviews, and content analysis.
- 3.1 Study design and procedure: Participants designed a culture center on an island site and produced a floorplan, interior perspective, and facade material sample without digitally editing generated images.
- 3.1 Study design and procedure: Three sessions gave participants approximately 1 hour 15 minutes to 1 hour 25 minutes for individual work while permitting discussion.
- 3.2 Data collection: Researchers administered the Creativity Support Index, conducted semi-structured group interviews, and analyzed interview commentary using content analysis.
- 3.2 Data collection: The participant table records demographics, architecture-study experience, session, assigned tool, and prior image-generation experience.
- The study used Midjourney, DALL-E 2, and Stable Diffusion, representing publicly accessible text-to-image generation tools.
- Seventeen architecture students participated, with varied study experience and limited prior exposure to image generators.
4 Results
Participants produced understandable concepts and completed the required visual outputs, while generated images enabled varied and sometimes adventurous architectural exploration but posed substantial challenges for floorplans and materials.
- All participants delivered understandable floorplans, material samples, and interior views, with sufficient time for ideation and exploration.
- Winning concepts combined strong presence and shape language with wood and organic forms that are difficult and time-consuming to model conventionally.
- Participants explored abstract, ornate, organic, site-specific, and unconventional material approaches, although the meaning of some decorative choices remained unresolved.
- Floorplans were rarely conventional black-and-white drawings and often contained colored, three-dimensional, or nonsensical layouts.
- Material generation was also difficult because the systems often produced ineffective facades, forcing participants to use simpler material images or exterior perspectives.
4.2 Creativity support
Creativity-support scores did not differ significantly among DALL-E, Midjourney, and Stable Diffusion, while participants’ experiences included varied and unexpected uses of generated images.
- CSI scores showed no significant differences among DALL-E, Midjourney, and Stable Diffusion: DE 75.2, MJ 70.9, and SD 73.7.The Kruskal-Wallis test yielded χ2(2) = 1.04 and p = 0.59.
- Participants used image generators unexpectedly, including abstract floorplans, ornate facades, honeycomb materials, site cues, strawberry materials, and experimental alternatives to unusable floorplans.
4.3 Prompts
Participants used descriptive prompts and iterative sequences to develop architectural ideas, generally staying close to the design brief while occasionally experimenting with alternative language and concepts.
- 4.3 Prompts: Participants wrote 588 prompts across the sessions, averaging 39.2 prompts per participant and 15.4 tokens per prompt.
- 4.3 Prompts: Prompts were generally descriptive and used terms intended to specify architectural outcomes.
- 4.3 Prompts: Examples addressed floorplans, elevations, materials, styles, perspectives, and cultural-center concepts across the three generators.
- 4.3.1 Prompt sequences: Participants grouped related prompts into sequences and typically started new sequences when results were unsatisfactory or when beginning another task.
- 4.3.1 Prompt sequences: Each sequence began with an initial prompt that participants iteratively extended with keywords to improve generated images.
- 4.3.1 Prompt sequences: Sequences reached 24 prompts, averaging 8.4 prompts, indicating sustained engagement with the image-generation tool during the design task.The standard deviation of sequence length was 5.9 prompts.
- 4.3.1 Prompt sequences: Most ideas generated fewer than four prompts before participants moved to a new prompt sequence.
- 4.3.2 Prompt language: Participants largely adopted terminology from the design brief, especially “floorplan” and “facade,” while only a few experimented with synonyms or other terms.
4.4 Qualitative insights
Participants found text-to-image tools easy and inspiring, but their usefulness depended on active prompting, sensemaking, and design constraints. The tools encouraged exploration and serendipitous ideas while also shaping how participants approached architectural ideation.
- Synthetic images in the architectural ideation process: Image generators supported imaginative, flow-like exploration and the serendipitous discovery of design ideas.Participants described the tools as easy, fun, inspiring, and an extension of their imagination; “Exploration” was the most important CSI subfactor.
- Synthetic images in the architectural ideation process: Participants valued generated images that could represent their own design concepts more precisely than existing inspirational images.
- Synthetic images in the architectural ideation process: Image generation was not a neutral substitute for architectural thinking because designers had to guide the system and assess its relevance to broader design goals.Participants noted that architecture involves more than appearance and includes meeting stakeholder needs.
- Randomness and sensemaking: Participants managed randomness through alternative interpretations and experimentation, sometimes allowing unexpected outputs to stimulate new ideas.Some participants deliberately tried to let the system go wrong and then evaluated whether the results made sense.
- Learning strategies: Effective prompting required adapting architectural vocabulary, keyword order, prompt length, and iterative refinement to the generators’ behavior.Participants sometimes replaced familiar terms such as “floorplan” with alternatives such as “plan drawing” or “plan view.”
- Learning strategies: Participants proposed curated recommendations, localized image understanding, and explicit architectural constraints such as size and number of floors.They also saw potential in integrating generative features with traditional computer-aided design software.
4.5 Unexpected results and prompting challenges
Participants encountered failures when generators could not represent local context, remove unwanted content, or produce feasible architectural elements. Prompt refinement improved image quality but could also constrain exploration and lead users away from their intended concepts.
- Unexpected results and prompting challenges: Local buildings, landmarks, and city features were not meaningfully recognized, causing participants to abandon location-specific prompts.
- Unexpected results and prompting challenges: Removing unwanted objects and text was difficult because few participants used negatively weighted prompts and prompt-based workarounds were limited.Literal references to natural structures and materials could also produce unexpected architectural forms.
- Unexpected results and prompting challenges: Generated materials and structures could be difficult to manufacture, structurally unsound, or infeasible for their intended purpose.
- Unexpected results and prompting challenges: Even skillful prompt refinement rarely guaranteed the intended result, and overuse of prompting could limit the creative process.Participants sometimes returned to earlier prompt versions after exploring unproductive variations.
5 Discussion
The study suggests that text-to-image generators can enrich architectural ideation when students use them reflectively and account for tool limitations, design constraints, and contextual creativity. It also identifies interface, representation, prompting, and educational challenges that should guide future tools and practices.
- 5.1 Ideation and creativity with AI: Text-to-image generators can support serendipitous idea discovery and an imaginative mindset when design constraints and reflective engagement are carefully considered.The authors frame effective use as dependent on understanding students’ motivations, abilities, and the affordances of chosen tools.
- 5.2.1 Considerations for the design of image generators: Prompt wording strongly affects generated results, making prompt writing a learned skill and a practical consideration for architectural design tools.The discussion notes that longer prompts typically produce higher-quality results, while prompt language requires substantial effort.
- 5.2.1 Considerations for the design of image generators: Architectural image generators should better support conventional floorplan generation, whose abstract nature and limited training data constrain current outputs.The authors identify dataset availability, labeling, curation, and context-specific architectural features as open development issues.
- 5.2.1 Considerations for the design of image generators: Spatial rather than vertically scrolling interfaces could help designers track idea development, explore design aspects, and zoom out across the ideation process.
- 5.2.2 Considerations for educators: Architectural education should begin with why image generation is useful, then teach effective use while preserving students’ reflective and imaginative agency.
- 5.3 Limitations: The study’s inexperienced participants and limited task setting mean that better instructions or extended tasks could produce better images and clarify current challenges.
6 Conclusions
The study examined how architecture students adopted text-to-image generation during early concept ideation. Its findings support using these tools as part of architectural design while attending to their limitations and educational implications.
- 6 Conclusions: A laboratory study with 17 architecture students used image generation to examine adoption during early architectural concept ideation.