Delirious Facade at DesignTO
Delirious Facade at DesignTO
Jan 24–Feb 28 2025
Window Exhibit
as a part of DesignTO at
Swipe Design Books + Objects
401 Richmond St W B04
Toronto, ON
exterior 24h / interior 11-6 tu-sa
DESIGN IN DELIRIUM
a research project exploring the integration of image processing artificial intelligence (AI) as a creative tool in architectural facade design.
DELIRIOUS FACADE
Article:
Vivian Lee
Model:
Jessica Babe
Sergej Mileski
Line Drawings:
Eisa Hayashi
Zhang Meiyun
Elmer Lyu
LAMAS Team:
Vivian Lee
James Macgillivray
Kara Verbeek
William Guinane
Madeleine Cheung
Our investigation is twofold:
1. focusing on the formal outcomes of AI-human collaboration
2. documenting the division of labor in the design process
The primary objective is to gain insights into the transformative impact of recent AI advancements on architectural design and the implications for architectural production.
Our research documents the impact of AI on architectural design by focusing on the dynamics of AI-human collaboration. We closely examine and evaluate design intentionality, prompt engineering, and the synthesis of ideas into a cohesive architectural model. This AI infused design process sheds light on new pathways of labor in architectural production, fundamentally reshaping the designer’s relationship to form generation. In addressing these inquiries, our work provides valuable insights into the effective use of AI as a design tool.
BACKGROUND
This project builds upon prior research in AI-facilitated architectural design, where our office employed Google Deep Dream, a GANs-based image processing application, to generate architectural façade design. In addition, our earlier research methods were deployed across two research studios for graduate and undergraduate architecture students, offering insights into both designing with AI and the pedagogical implications of its use.
More recent advancements in AI applications, exemplified by technologies like Midjourney and DALL-E, have revolutionized the speed and ease of image generation, fundamentally transforming the design process. While in our earlier research we had to critically input two parent images to yield a hybrid result (see figures 3 and 4) - the new Diffusion Model of AI image generators can respond with complete building images based on simple text prompts. Furthermore, the image results of these queries often ignore the actual input precedents altogether and impose a stylistic predilection of its own.
This newfound capacity for rapid image generation raises concerns in architectural design production. The facility with which a seemingly fully formed building can be conjured is a distraction to inexperienced designers. There is a risk that these powerful AI tools might invite us to neglect foundational design considerations critical to our discipline. Consequently, Delirious Façade is not only a design-focused endeavor but also one that scrutinizes architectural labor and production, recognizing a shift in the skill sets essential for architects.
PROJECT BRIEF
FIG. 3 Using GANs software, Syracuse University student Genevieve Dominiak “face-swapped” Renzo Piano’s Centre Pompidou with Ca’ d’Oro by Giovanni Bon.
While earlier research worked mostly with rasterized images, this project demanded the creation of a fully resolved 1:20 scale physical model of the proposed façade. The ambition to transform raster 2D imagery to a coherent 3D building façade exposed the mechanisms for designing with AI, in this case Midjourney. We steered the AI to clarify window details, railing patterns, while supplementing its freeform imaginaries with our architect-trained expertise in material assemblies and historical understanding of facade compositional strategies. The result is a remarkable physical model and a series of diagrams detailing the human-robot design process.
The design of buildings encompasses many aspects of invention and coordination. For the purposes of this project, we’ve narrowed our focus exclusively to façade design. This framed engagement of design aligns with the astonishing capabilities of recent AI technology, generating ambiguous images that inspire formal and compositional innovation.
Much like our earlier research inquiries with AI, Delirious Façade begins by hybridizing Toronto facades – in this case we chose one hybridized result - an iconic Uno Prii apartment blended with a vernacular turn-of-the-century mixed-use building on College Street. The challenge was to interpret and densify this hybrid image into a cohesive five-story façade design.
FIG. 4 Uno Prii’s Apartment Building and a vernacular Toronto commerical building is hybridized. The middle image shows a GANs result versus the bottom image demonstrates the latest diffusion model output.
DESIGN PROCESS
When presented with a multitude of variegated imagery, how can we discern meaning within the delirium? Amid this onslaught of ideas, it becomes of utmost importance to underscore the significance in establishing a thesis, a thread of purpose to the design exercise.
For the starting hybrid image, we did trials with both the GANs and Diffusion Model AIs. In figure 4, one can see the difference; where the Diffusion Model provides a clearly resolved (if not mundane) result of the façade blend, the original GANs version is much more akin to a surrealist painting. In reviewing these two versions, we (humans) agreed that Midjourney’s version is a foregone conclusion, but that the GANs’ image offers more formal potential for interpretation.
As a test, we took the chosen base hybrid image and asked Midjourney to describe the .jpg into text. We then fed the descriptive text back into Midjourney to translate into images. The results (see in figure 5) were entirely incoherent and not reversible in loop processing. In fact, performing a second sequencing of the same procedure yields a whole other assemblage of architectural styles, framed through a three-quarter perspective.
Our lack of ability to reverse the processes of the AI consistently reflects the fact that AI in image processing is essentially not auditable. These apps are defacto black box processors and only in very rare cases can be understood as “step by step” processes. This puts them necessarily at odds with the iterative and process based models of architectural design and pedagogy.
FIG. 6 The results of a second iteration of image to text back to image through Midjourney. Each iteration yields foreign results from the previous input images.
FIG. 5 The results of inputting the descriptive text generated by Midjourney of our base image, then fed back into Midjourney to generate images.
Another test is to imagine how other famous architects might interpret the base image. Inputting the base image with a text prompt to perform “architecture models in the style of Aldo Rossi, Antoni Gaudi, Carlos Scarpa, Mies Van Der Rohe, SANAA, Bjarke Ingles”– it was surprising that the AI rendered all six architects with similar results (see figure 7.) While AI Gaudi’s version might have the most arcuated bio-morphic form true to this buildings, AI Mies’ design was surprisingly curvilinear with melted columnar supports.
Our experimentation led us to examine architectural prompt engineering in more detail. We (humans) made an interpretation of the base image through Rhino and 3ds Max digital modeling. This new base image – both a screenshot of the rendered 3D digital model and a photograph of the 3D printed physical model – where used as the new base image for stylistic interpretation (see figure 8.) Where AI proved most insightful was when we fed both the digital and physical 3D images back into Midjourney, specifically requesting a response on “articulated doors and windows.” This allowed us to harness AI to help interpret building features present in the base image
In fact, where the “in the style of ” is most successful is when we insert our base image of the photograph of the 3D model AND another representation of the same kind (in this case a photograph of a 3D model by Caruso St. John) and were able to achieve interesting Caruso St. John style results when prompted “architectural model, pastel, flowing, flat, clean.” (see figure 9.).
FIG. 7 A series of Midjourney tests to design “in the style of” famous architects. Text and a base image was used for each of these series. A evidenced in this array, the source image geometry and representational modes tend to sway the outputs more visibly than the text prompt itself.
In summary, three prompting lessons can be learned from this experiment.
First, providing the AI with more than one image of the intended result enhances its ability to triangulate and refine outputs.
Second, text prompts should encompass various descriptors, including representation type, colors, formal attributes, and emotional qualities, as more adjectives tend to generate detailed images.
Lastly, query for specific architectural elements such as door frames, windows, and railings. This enables the AI to zoom into these areas and envision architectural details and materiality.
FIG. 8 Both the 3D print and screenshot render of the digital model were fed into Midjourney along with specific text prompts to articulate windows and doors.
FIG. 9 Inserting more than one image along with text prompt into Midjourney helps to steer the outputs. Here a Caruso St John Architects model was used alongside more descriptive adjectives in the text prompt.
LESSONS LEARNED
Throughout this design process, we found it imperative to consistently re-align ourselves with our core objective. The goal was to create a hybrid façade from two disparate architectural typologies. Each precedent building had its own logic, ornamentation, and occupiable spaces. We were tasked to transform a raster image of this hybrid façade into a physically accountable model that demonstrated material and spatial logic.
In addition to some of the prompt tips provided in the previous section, it is also important to realize that there is no “perfect prompt” capable of steering AI toward a precise intentional output. Instead, AI serves as a tool for generating formal ideas, with its immediacy and wealth of solutions aiding designers in making informed decisions.
Designing by attrition still means that the designer will have to synthesize the options presented into one cohesive project.
FIG. 11 Detail photo of model exterior
FIG. 10 Detail of model interior
The “will to form,” which can be daunting for early designers, is readily encouraged through AI assistance. Designers can behave more like editors in future, assessing the various outputs for fitness to the site, program, design concept and overall context. This tool has the potential to empower community members who may lack the skills to articulate their design preferences, fostering inclusivity in the design process.
Furthermore, designing with AI has transformed the nature of discussions and critiques among designers. Exposing students and professionals to a multitude of design options, none exclusively crafted by a single human, encourages productive critique. In our research, we observed that presenters felt less personally attached to the production outputs, fostering a willingness to engage in judgment and high-level conceptual conversations. This aspect of AI-facilitated design has the potential to enhance studio culture and advance a new model for studio-based iteration and dialogue.
The setup of this research project – with a focus on large scale models and material definition – provided a valuable experience. Designers had to further research (using traditional resources) the realities of building systems, material assemblies, and historical architectural ornament usages for the resolution of the physical model. This undertaking highlighted the potential for establishing a disciplinary approach whereby AI expedites the Schematic Design phase of a project, enabling a broader exploration of comprehensive real-world building-related topics.
Ultimately, the lessons gleaned from this collaborative design process with AI underscore the enduring presence of the trained architect’s agency throughout the design journey. AI in visual design can be likened to a surrealist technique, an expedient method for breaking away from conscious composition and conventional tropes. AI grants designers the freedom to explore and discover novel arrangement in forms - but in the end, the architect’s design vision must rein the AI’s musings into a harmonious whole.
