Think AI-side the Box! Exploring the Usability of Text-to-Image Generators for Architecture Students

Abstract

This study examines how architecture students use generative AI image generating models for architectural design. A workshop was conducted with 25 participants to create designs using three state-of-the-art generative diffusion models and BIM or 3D modeling software. Results showed that the participants found the image-generating models useful for the preliminary design stages but had difficulty when the design advanced because the models did not perform as they expected. Finally, the study shows areas for improvement that merit further research. The paper provides empirical evidence on how generative diffusion models are used in an architectural context and contributes to the field of digital design.

Full text

Think AI-side the Box!
Exploring the Usability of Text-to-Image Generators for Architecture Students
Jonathan Dortheimer1, Gerhard Schubert2, Agata Dalach 3, Lielle Brenner4
Nikolas Martelaro5
1,4 Ariel University, 2,3 Technical University of Munich, 5 Carnegie Mellon University
1jonathand@ariel.ac.il 2,3{schubert|agata.dalach}@tum.de
4liellejo.brenner@msmail.ariel.ac.il 5nikmart@cmu.edu
This study examines how architecture students use generative AI image generating
models for architectural design. A workshop was conducted with 25 participants to create
designs using three state-of-the-art generative diffusion models and BIM or 3D modeling
software. Results showed that the participants found the image-generating models useful
for the preliminary design stages but had difficulty when the design advanced because the
models did not perform as they expected. Finally, the study shows areas for improvement
that merit further research. The paper provides empirical evidence on how generative
diffusion models are used in an architectural context and contributes to the field of digital
design.
Keywords: Machine Learning, Diffusion Models, Design Process, Computational
Creativity.
INTRODUCTION
Recent advances in generative machine learning
models have demonstrated the potential to
revolutionize the field of architecture by introducing
a new kind of creative behavior enabled by Artificial
Intelligence (AI) tools. For example, Generative
Adversarial Networks (GAN) models have already
been used in specific architectural tasks such as plan
generation (Chaillou, 2020; Zheng et al., 2020),
concept image generation (Silvestre et al., 2016;
Eroğlu & Gül, 2022) and urban planning (Boim et al.,
2022; Zhong et al., 2022). Recently, new image
generation diffusion models have demonstrated an
extraordinary capability to produce novel and
creative architectural illustrations, including
architectural designs from text descriptions and
images (Ramesh et al., 2021).
It remains unclear how architects can use the
diffusion models in architectural practice, which
models are more appropriate, and how the use of
the models affects the design process. To address
this knowledge gap, we conducted an experimental
workshop with 25 architecture students to design
using three state-of-the-art diffusion models (Dall-E,
MidJourney, and Stable Diffusion) and BIM/3D
modeling software (Revit). This study’s research
question is “How could image generating models be
used in the context of the architecture studio?” This
paper contributes empirical evidence on how
architects use generative diffusion models to
produce designs and identifies some opportunities
and challenges using these models in practice.
Volume 2 – Digital Design Reconsidered – eCAADe 41 | 567

AI IN ARCHITECTURE
Architecture constantly evolves due to aesthetic
trends and the development of new design tools,
such as computer-aided architectural software.
These modern tools had a significant impact on the
architectural process. Early predictions for computer
use in architecture included task automation,
alternative work methods, and machine partnership
in design evolution (Negroponte, 1972). Drafting
software and BIM address the first two visions, while
AI tackles the third.
Antoine Picon (2020) wrote: "Artificial
intelligence is about to reshape the architectural
discipline." In his paper, he pointed out that AI
technologies could improve efficiency, help design
new things, relieve humans from performing
repetitive labor, and support them in creative tasks.
Several approaches address the challenge of AI
supporting architects' design processes. While early
research, such as Simon (1969), viewed design as a
search in the solution space, later studies considered
it an exploration (Come et al., 1994). Design
problems are often ill-defined, requiring exploration
of the problem space (Maher et al., 1996). AI can
augment design by extending solution exploration,
similar to how multiple human designers approach
the same problem (Dortheimer, 2022).
Generative AI
Breakthrough research in machine learning (ML) has
demonstrated a high ability to create new images
using unsupervised learning (Goodfellow et al.,
2014). Different models have produced novel
images based on a given text prompt, known as
reverse image caption (Mansimov et al., 2016).
Diffusion models further improved the text-to-
image transition, transforming noise into realistic
pictures (Saharia et al., 2022). State-of-the-art models
include Stable Diffusion (Rombach et al., 2021),
Imagen (Saharia et al., 2022), DALL-E 2 (Ramesh et al.,
2021), and MidJourney.
Several approaches have been made to study the
use of such technologies in architecture. One of the
first approaches was using convolutional neural
networks to generate new architectural pictures
(Silvestre et al., 2016). Later, DALL-Pytorch was
trained to generate new drawings and plans suitable
for architecture (Bolojan, Vermisso & Yousif, 2022). In
another study, Pix2pixHD was trained using floor
plan datasets (Zheng et al., 2020). The researchers
were able to generate realistic floor plans in the
given style, though not always with functional logic.
Other studies involved several GANs sequentially to
floor plan generation (Chaillou, 2020).
Another possible application is using text-to-
image generators for architectural building images
(Zhu et al., 2018; Eroğlu & Gül, 2022). Moreover, GANs
may also find their purpose in urban planning.
Previous research shows that AI can learn urban
patterns and generate resembling patterns (Boim et
al., 2022).
However, text-to-image generators require well-
crafted text prompts to produce satisfactory results,
leading to the development of "prompt
engineering" (Oppenlaender, 2022). Still, using text
alone for architectural image generation has
limitations due to the reductive nature of textual
descriptions and the need for explicit architectural
training in general-purpose models (Bolojan et al.,
2022). Researchers have explored less specific
prompts, GPT-3 keyword suggestions (Liu et al.,
2022), and generators that incorporate schematic
drawings alongside text prompts (Gafni et al., 2022).
Machine and Human Creativity
Creativity could be understood as creating original
content (Boden, 2004) but is typically perceived and
evaluated subjectively (Mrosla, Koch & von Both,
2019). According to Boden (2004), there are three
types of creativity: combinational, exploratory, and
transformational. Creativity in artistic forms was
already targeted by computational systems, while
568 | eCAADe 41 – Volume 2 – Digital Design Reconsidered

creative problem-solving received less attention in
research (Oltețeanu, 2020).
Several studies explored how humans and AI
models co-create design using images as inspiration.
While AI-generated images may be perceived as less
creative, combining human and machine creativity
can enhance innovation (Huang et al., 2021),
Research has explored workflow strategies (Yousif &
Vermisso, 2022), community collaboration (Epstein
et al., 2022), and direct cooperation between
designers and AI (Gmeiner et al., 2023).
Other studies claim that AI can benefit low-
performing teams (Zhang et al., 2021). The
researchers suggested that the communication
issues between humans and AI could be solved by
referring to human-human cooperation methods
(Gmeiner et al., 2023). Nevertheless, it can be
assumed that various design problems will require
different AI-human collaboration methods and
workflows.
METHOD
To answer the study’s research question, “How could
image-generating models be used in the context of
the architecture studio?”, the research team hosted
the three-day architecture student workshop aimed
to expose students to diffusion models and allow
them to use the models through an architectural
design task to produce a small building using any
design tools they choose. The workshop provided a
valuable opportunity for the research team to
conduct an exploratory study to learn how the
students use these models with a considerable
design task.
Participants
The workshop was open to architecture students in
their third to fifth year of study. A total of 25 students
participated, with 18 females and seven males: two
were in their fifth year, one was in their fourth year,
and the remaining 17 were in their third year. To
minimize the pressure on the participants, the
workshop was designed to ensure that the student's
scores were not dependent on their design
performance. Therefore, they were awarded one
credit for their participation, whether they produced
any design using AI tools or not.
Workshop Structure
In this three-day workshop, participants engaged in
a two-phase process that combined lectures and
exercises to explore the possibilities, limits, and
framework conditions of AI models in architectural
design.
Phase 1, the experimental exploration day, was
based on Italo Calvino's "The Invisible Cities."
Participants selected a city from the book, imagined
it, and then visualized it using various AI image
generators, such as Stable diffusion, Dall-E, and
MidJourney. This introductory exercise aimed to
develop the semantics for image generation
through a design-based approach. Participants were
asked to reflect on their experiences, successes, and
challenges while creating the images, comparing
their imaginary visions with the AI-generated
outcomes, and evaluating the tools and workflows
used.
In Phase 2, participants were tasked with employing
AI models to generate a three-dimensional building.
This phase focused on the autonomous and self-
directed implementation of the models for inventive
architectural design. The design task involved
creating a multicultural community building with
two floors, 200 square meters in size, and including
stairs to connect the floors. The building supposed
to be designed to celebrate the diverse cultures of
the local population. This task aimed to make
participants deal with the complexities of
transforming 2D images into 3D models,
incorporating all the necessary systems and
components for a functional and culturally inclusive
space.
Volume 2 – Digital Design Reconsidered – eCAADe 41 | 569

Data Collection
The research data includes final presentation slides
and recordings of the participants' computer screens
using Zoom software (see Figure 1). The recordings
documented the participants' use while working
with the AI models and CAAD software. When the
participants finished the design, they saved the
recording file and provided it to us. According to the
Internal Review Board approval, the students were
free to participate in any part of the study without
affecting their grades.
Data Analysis
The research team viewed, analyzed, and coded the
recordings. The coding scheme includes the
software used (Dall-E, MidJourney, Stable Diffusion,
Revit, and SketchUp), the text prompts the
participants provided to the AI models, and the
prompt kind (text, image, variation, and image
extension). Special notice was provided to the kind
of architectural medium the participant requested,
such as a plan, section, elevation, or 3D model, and
whether the participant used the generated result.
When sequential prompts were provided, the team
identified the changes between the prompts, the
architectural medium requested, the generated
output, the changes to the prompt, and if the
participant used the output.
RESULTS
The research team concluded the workshop
according to plan, having collected 25 student
presentations and 16 video recordings with an
average duration of 4:02 hours. Three research team
members then conducted a qualitative analysis of
the video recordings, noting 645 instances of CAD
software usage with an average of 40.31 per
recording (SD 16.62).
Image Generator Usage and Efficiency
The utilization of AI models was documented in
Table 1, with a mean of 36.62 (SD 17.03) instances of
usage. MidJourney was the most employed model,
with an average of 13.93 (SD 11.35) instances,
followed by Dalle-E (M = 13.00, SD = 12.12), Stable
Diffusion (M = 6.43, SD = 7.42), and Lexica (M = 2.25,
SD = 2.04). Additionally, four participants utilized
ChatGPT to facilitate discourse and articulate the
text prompts.
The average success rate of AI models was
calculated as the average number of successful
usage instances divided by the total number of
usage instances (see examples in Figures 2, 3, 4).
MidJourney had the highest success rate at 39.85%
(SD 20.68%), while Dall-E, Stable Diffusion, and
Lexica had success rates of 22.88%, 23.93%, and
21.70%, respectively. These results suggest that
MidJourney outperformed the other models
Figure 1
Data collection and
analysis
570 | eCAADe 41 – Volume 2 – Digital Design Reconsidered

regarding participants' satisfaction with the
outcomes.
Architectural Prompt Anatomy
This study additionally provides an analysis of the
terms used to generate architectural images.
Specifically, three categories of terms were
identified: object properties, situation description,
and image properties. Object properties encompass
building usage, materials, building details, building
size, architectural style, and building form. Examples
of object properties include building usage (e.g.,
community center, art center, entrance lobby),
materials (e.g., stone, concrete, glass, iron, colors),
building details (e.g., Arabic ornaments, stone
carvings, inner courtyard, vertical shutters), building
size (e.g., two stories, 200 square meters, eight
meters high), architectural style (e.g., modern,
architect name, futuristic, crystal palace, Arabic), and
building form (e.g., teardrop, organic shape).
Situation description includes building
placement (e.g., underground, in the field), building
relationship (e.g., facing each other, single building,
two buildings, next to a road), architectural scale
(e.g., 1:00, 1:500), and geographic location (e.g.,
Singapore, New York, Melbourne).
Image properties encompass architectural medium
kinds and photograph properties. Examples of
image properties include architectural medium
kinds (e.g., 3D model, plan, section, elevation, sketch,
render image) and photograph properties (e.g., wide
angle, axonometric, bird view, Lumion render, ultra-
HD). The research team observed several intriguing
applications of text prompts. One participant
employed the phrase "correlated plan and section"
to generate 3D renders with plans and sections,
which was beneficial. Additionally, another
PCPs
Counted Instances Success Rate
Mid
Journey
Dall-E
Stable
Diffusion
Lexica
Mid
Journey
Dall-E
Stable
Diffusion
Lexica
1 16 4 5 2 56.25% 50.00% 40.00% 0.00%
2 2 17 23 6 50.00% 52.94% 52.17% 16.67%
3 21 15 1 1 42.86% 40.00% 0.00% 0.00%
4 15 11 3 2 40.00% 9.09% 33.33% 0.00%
5 10 3 7 4 10.00% 0.00% 0.00% 25.00%
7 5 1 0.00% 0.00%
11 6 3 1 1 83.33% 66.67% 0.00% 0.00%
12 26 31 3 19.23% 12.90% 0.00%
13 23 6 9 1 56.52% 0.00% 22.22% 100.00%
14 5 23 4 40.00% 17.39% 50.00%
16 19 1 3 2 42.11% 0.00% 33.33% 50.00%
21 17 23 7 47.06% 21.74% 57.14%
22 25 15 3 16.00% 26.67% 0.00%
23 20 45.00%
26 15 39 12 3 33.33% 23.08% 8.33% 33.33%
27 43 6 1 37.21% 33.33% 0.00% 0.00%
Avg. 13.94 13.00 6.44 2.25 39.85% 22.88% 23.93% 21.70%
SD 11.35 12.12 7.43 2.05 20.68% 21.42% 20.21% 31.17%
Table 1
Summary of model
usage instance and
computed success
rate for each
participant
Volume 2 – Digital Design Reconsidered – eCAADe 41 | 571

participant combined the styles of two architects
into a single edifice by utilizing their respective
names.
The qualitative analysis of the 25 student
presentations revealed a greater understanding of
the participants' perspectives on utilizing the
models. The presentations also enabled the research
team to ascertain which generated images were
eventually employed by the participants. This
analysis enabled the research team to discern the
design activities in which the models were utilized
and the efficacy of their implementation (see Figure
2). All the participants employed the AI models to
explore design ideas based on the design brief text.
The participants were free to select any AI models,
with some opting for a single model and others
utilizing several models. To further develop the
design idea, some participants used the `variations'
feature or uploaded images with or without a text
prompt to generate new variations.
In the later stages of the design process, some
participants attempted to use AI to generate two-
dimensional architectural drawings of plans,
sections, and elevations. However, these efforts were
mostly unsuccessful, as the generated images
tended to vary from the original architectural image
rather than providing a two-dimensional
interpretation. Even when the participants
successfully used the generated image, the results
were only loosely related to the original images. The
visual search was over at a certain point, and the
participants employed modeling software (e.g.,
Revit) to produce a 3D model from one or more
images.
Initially, the participants compared the work-in-
progress 3D model to the AI images. However, this
became less frequent as the modeling process
progressed. Interestingly, most participants did not
employ the AI models to explore the design during
the modeling phase further.
Participants generally tended to remain faithful
to the selected AI images, which we argue is a design
fixation (see Figure 3). The clearer and more precise
the images were, the more the participants
replicated the design in the image rather than
dealing with the complexity associated with
transitioning it to 3D.
One participant was observed to provide a 3D
section screenshot image to Dall-E and requested
that the section be developed in a specific known
architect's style. Following the development of a
model, some participants employed AI models to
enhance the renders of the 3D building model. This
entailed modifications to the weather, expansion of
the image, and the incorporation of local context.
Additionally, one participant utilized the generated
images as artistic components to decorate an
interior space render. One participant chose to pass
from using the AI and created their design
independently instead. The participant explained
that she was dissatisfied with all the outputs and,
thus, rejected its involvement in the creative process.
Figure 2
A typical design
workflow of one of
the participants
572 | eCAADe 41 – Volume 2 – Digital Design Reconsidered

DISCUSSION
This study of how architecture students use image-
generating models for design revealed that the
models are best used in the early ideation phase.
Additionally, examining text prompts used in the
process uncovered distinct patterns that
contributed to creating architectural designs.
Interestingly, even though students used the models
with similar frequency, they repurposed the outputs
from MidJourney significantly more (40% of the time
compared to 23% for the rest). These findings help to
further our understanding of the role of image-
generating models in the architectural design
process.
Our results have demonstrated the potential of
generative models in computational creativity,
specifically in architectural design. Utilizing
combinatorial creativity, as suggested by Boden
(2004), our results indicate that generative models
can be used to search for ideas, precedents, and
inspiration, as well as to merge concepts, such as two
architectural styles, to produce creative outcomes
(see Figure 4). This suggests that generative models
can be used to explore various novel design
solutions, thereby enabling architects to leverage
computational creativity to explore new aesthetic
possibilities.
Some results suggest that the use of AI models
to create design images can be problematic, as it can
lead to design fixation. Specifically, when
participants used MidJourney, which output
illustrations were more "artistic," the participants
tended to remain more consistent with the
generated images, copying the features precisely
and producing a result very close to the picture (see
Figure 3). In contrast, participants that used more
models that made more realistic pictures, such as
Dall-E and Stable Diffusion, demonstrated more
creativity in creating architecture, as the design in
the generated images was less clear. This suggests
that participants may have become "fixated" on
MidJourney's beautiful illustrations, being less
critical, and did not make the necessary adjustments.
These findings have important implications for using
AI models in the design process, as they suggest that
design fixation can occur.
Furthermore, the utilization of the models in the
advanced phases of the design process has been
limited in their efficacy. This was evident when
participants attempted to use the models to develop
designs. The participants tried to provide an image
of the designs they wanted to develop with text
prompts to improve the design or transform
between architectural blueprints (e.g., plans,
sections, or elevations), yet the results were
frustrating. This is likely due to the random nature of
the diffusion models, which can generate various
images, but at the expense of continuity. These
results hinder the application of these models for
design iterations when the goal is to advance the
design.
Figure 3
Various examples
of image to 3D
conversion. On the
left, MidJourney
generated images,
on the right, the
produced BIM
models using Revit.
Volume 2 – Digital Design Reconsidered – eCAADe 41 | 573

In the context of generating blueprints, the models
failed to adhere to engineering logic. The blueprints
were unclear, spaces were not defined, openings
and doors were missing, and sometimes there was a
mix between plan and section views. Furthermore,
when multiple plans, sections, and facades were
included in a single image, no discernible
engineering logic connected them. These models
did not have the knowledge of architectural
principles, which was partly demonstrated by
previous works (Chaillou, 2020).
Nonetheless, the simplicity of using these
models for architectural expression has the potential
to revolutionize how clients communicate their
ideas. By allowing non-professionals to create new
and innovative images and designs, AI models can
provide a platform for the visual articulation of ideas
that have been difficult to express without graphical
skills. However, to effectively utilize this tool,
acquiring knowledge of architectural terminology,
as identified in this study, and gaining experience in
using the instrument is necessary. Additionally,
architects can use AI models to generate multiple
design alternatives to present to clients, thereby
facilitating the exploration of design directions.
Limitations
The limitations of our study stem from its exploratory
nature, focusing on architecture students in a studio
setting, which may not accurately represent the
experiences of professional architects, potentially
influencing creative fixation.
CONCLUSION
In conclusion, our study explored the potential of
image-generating AI models in architectural design
by examining their use in a studio context. We
analyzed model usage and architectural prompts
identifying their structural anatomy, benefits, and
limitations. Our findings show that AI models can
inspire creativity and innovation. However, they face
challenges like causing design fixation, having
limited design development capability, and
generating non-coherent blueprints. These
challenges highlight the need for further research to
improve AI models' effectiveness in architectural
design.
Future research should focus on reducing
creative fixation, enhancing AI models' applicability
to advance a design by better understanding human
intent, and improving architectural logic by
developing AI models that generate coherent plans
and corresponding sections. With these research
directions, we aim to contribute to the ongoing
discussion on AI's role in architecture and facilitate
more effective integration of image-generating
models in the architectural design process.
REFERENCES
Boden, M.A. (2003). The Creative Mind: Myths and
Mechanisms (2nd ed.). Routledge.
https://doi.org/10.4324/9780203508527https://
doi.org/10.4324/9780203508527
Boim, A., Dortheimer, J. and Sprecher, A. (2022) ‘A
Machine-Learning Approach to Urban Design
Figure 4
Example of
combinational
creativity using a
prompt that
merges Gaudi’s
with Zaha Hadid’s
style. Left images
by Dall-E, right
images by
MidJourney.
574 | eCAADe 41 – Volume 2 – Digital Design Reconsidered

Interventions In Non-Planned Settlements’, in.
CAADRIA 2022: Post-Carbon, Sydney, Australia,
pp. 223–232. Available at:
https://doi.org/10.52842/conf.caadria.2022.1.22
3.
Bolojan, D., Vermisso, E. and Yousif, S. (2022) ‘Is
Language All We Need? A Query Into
Architectural Semantics Using a Multimodal
Generative Workflow’, in. CAADRIA 2022: Post-
Carbon, Sydney, Australia, pp. 353–362.
Available at:
https://doi.org/10.52842/conf.caadria.2022.1.35
3.
Chaillou, S. (2020). ArchiGAN: Artificial Intelligence x
Architecture. In: Yuan, P.F., Xie, M., Leach, N.,
Yao, J., Wang, X. (eds) Architectural Intelligence.
Springer, Singapore.
https://doi.org/10.1007/978-981-15-6568-7_8
Dortheimer, J. (2022) ‘Collective Intelligence in
Design Crowdsourcing’, Mathematics, 10(4), p.
539. Available at:
https://doi.org/10.3390/math10040539.https://
doi.org/10.3390/math10040539.
Goodfellow, I.J. et al. (2014) ‘Generative Adversarial
Networks’. arXiv. Available at:
http://arxiv.org/abs/1406.2661 (Accessed: 19
March 2023). (Accessed: 19 March 2023).
Epstein, Z., Schroeder, H. and Newman, D. (2022)
‘When happy accidents spark creativity:
Bringing collaborative speculation to life with
generative AI’. arXiv. Available at:
http://arxiv.org/abs/2206.00533 (Accessed: 20
January 2023). (Accessed: 20 January 2023).
Eroğlu, R. and Gül, L.F. (2022) ‘Architectural Form
Explorations through Generative Adversarial
Networks - Predicting the potentials of
StyleGAN’, in. eCAADe 2022: Co-creating the
Future - Inclusion in and through Design,
Ghent, Belgium, pp. 575–582. Available at:
https://doi.org/10.52842/conf.ecaade.2022.2.57
5.
Gmeiner, F., Yang, H., Yao, L., Holstein, K. and
Martelaro, N., 2023. Exploring Challenges and
Opportunities to Support Designers in Learning
to Co-create with AI-based Manufacturing
Design Tools. arXiv preprint arXiv:2303.00192.
Huang, J. et al. (2021) ‘On GANs, NLP and
Architecture: Combining Human and Machine
Intelligences for the Generation and Evaluation
of Meaningful Designs’,
Technology|Architecture + Design, 5(2), pp.
207–224. Available at:
https://doi.org/10.1080/24751448.2021.196706
0.
Liu, V. et al. (2022) ‘3DALL-E: Integrating Text-to-
Image AI in 3D Design Workflows’. arXiv.
Available at: http://arxiv.org/abs/2210.11603
(Accessed: 28 February 2023). (Accessed: 28
February 2023).
Maher, M.L., Poon, J. and Boulanger, S. (1996)
‘Formalising Design Exploration as Co-
Evolution’, in J.S. Gero and F. Sudweeks (eds)
Advances in Formal Design Methods for CAD.
Boston, MA: Springer US, pp. 3–30. Available at:
https://doi.org/10.1007/978-0-387-34925-1_1. .
Mansimov, E. et al. (2016) ‘Generating Images from
Captions with Attention’. arXiv. Available at:
http://arxiv.org/abs/1511.02793 (Accessed: 20
January 2023). (Accessed: 20 January 2023).
Mrosla, L. and Koch, V. (2019) ‘Quo vadis AI in
Architecture?’ Available at:
https://doi.org/10.5151/proceedings-
ecaadesigradi2019_302. .
Negroponte, N. (1972) The Architecture Machine:
Toward a more human environment. 2. print.
Cambridge, Mass.: MIT Pr.
Oltețeanu, A.-M. (2020) Cognition and the Creative
Machine: Cognitive AI for Creative Problem
Solving. Cham: Springer International
Publishing. Available at:
https://doi.org/10.1007/978-3-030-30322-8
Oppenlaender, J. (2022) ‘A Taxonomy of Prompt
Modifiers for Text-To-Image Generation’. arXiv.
Volume 2 – Digital Design Reconsidered – eCAADe 41 | 575

Available at: http://arxiv.org/abs/2204.13988
(Accessed: 20 January 2023). (Accessed: 20
January 2023).
Picon, A. (2020) ‘What About Humans? Artificial
Intelligence in Architecture’, in P.F. Yuan et al.
(eds) Architectural Intelligence. Singapore:
Springer Nature Singapore, pp. 15–29. Available
at: https://doi.org/10.1007/978-981-15-6568-
7_2.
Ramesh, A. et al. (2021) ‘Zero-Shot Text-to-Image
Generation’. arXiv. Available at:
http://arxiv.org/abs/2102.12092 (Accessed: 20
January 2023).
Rombach, R. et al. (2022) ‘High-Resolution Image
Synthesis with Latent Diffusion Models’, in 2022
IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR). 2022 IEEE/CVF
Conference on Computer Vision and Pattern
Recognition (CVPR), New Orleans, LA, USA: IEEE,
pp. 10674–10685. Available at:
https://doi.org/10.1109/CVPR52688.2022.01042
Saharia, C. et al. (2022) ‘Photorealistic Text-to-Image
Diffusion Models with Deep Language
Understanding’. arXiv. Available at:
http://arxiv.org/abs/2205.11487 (Accessed: 14
February 2023). (Accessed: 14 February 2023).
Silvestre, J., Ikeda, Y. and Guéna, F., 2016. Artificial
imagination of architecture with deep
convolutional neural network. Living Systems
and Micro-utopias: Towards Continuous
Designing, Melbourne; The Association for
Computer-Aided Architectural Design Research
in Asia (CAADRIA): Hong Kong, China, pp.881-
890.
Simon, H.A., 1969. The sciences of the articial
Cambridge, MA: MIT Press
Yousif, S., Vermisso, E. (2022), Towards AI-Assisted
Design Workflow for an expanded design space.
Zhang, G. et al. (2021) ‘A cautionary tale about the
impact of AI on human design teams’, Design
Studies, 72, p. 100990. Available at:
https://doi.org/10.1016/j.destud.2021.100990
Zheng, H. et al. (2020) ‘Apartment Floor Plans
Generation via Generative Adversarial
Networks’, in. CAADRIA 2020: RE:Anthropocene,
Bangkok, Thailand, pp. 599–608. Available at:
https://doi.org/10.52842/conf.caadria.2020.2.59
9..
Zhu, J.-Y. et al. (2018) ‘Toward Multimodal Image-to-
Image Translation’. arXiv. Available at:
http://arxiv.org/abs/1711.11586 (Accessed: 1
February 2023). (Accessed: 1 February 2023).
576 | eCAADe 41 – Volume 2 – Digital Design Reconsidered