Preface
In recent years, Artificial Intelligence (AI) has made significant inroads into the creative industries, notably in fields such as architecture design. The advent of AI-driven image creation tools and advanced language models has opened new frontiers in ideation, enabling a unique synergy between human creativity and generative AI imaging. This independent study aims to delve deep into the potential impacts of AI on the discipline of architecture and its integration into the design process. Three architecture students from the University of Cincinnati, Alex, Jackman, and Nathan, have embarked on a comprehensive investigation, utilizing cutting-edge AI tools to enhance their design capabilities and challenge and redefine traditional methodologies. Throughout their journey, three students have explored various dimensions of AI applications in architecture – from conceptualization and visualization to more complex aspects like representation and style. We invite you to explore the fruits of their experiments, which demonstrate their skillful application of AI in architecture and offer a glimpse into the future of architectural design, where human ingenuity and AI collaborate to create extraordinary works.
Ming Tang
AI in architecture design
by. Alex Hamilton, Jackman Land, Nathan Rosenberger
Year in review
Through various experiments and trials of multiple proprietary software, we have some findings and processes to aid students in how they navigate the new and ever-changing world of AI. The software we have primarily focused on are Open AI’s Midjourney and Chat GPT, Prome AI, and Adobe’s AI. Our goal in this semester of study was to identify how AI affects the design process, both currently and in the future. Our findings have also helped to identify whether or not AI can close the gap in design that a college education offers.
Our Experiment
Our primary experiment dove into whether or not a college education in architecture would yield better results when writing prompts for midjourney renderings. There is an in-depth report on this experiment, here is the summary of it. Three architecture students and three other people of varying ages and education each wrote 4 prompts aiming to achieve 4 different outcomes: an abstract piece of art, an outdoor rendering, an indoor rendering, and a con concept sketch of a building. The prompts would then be loaded into midjourney and they assembled randomly and given to volunteer survey subjects to see if they could pick which 3 images in each category were developed by the architecture students.
Our Findings
Here is what we found from this experiment: while the complexity and length of the prompt does not seem to have a drastic effect on the final production quality, you are able to achieve a much more targeted result by understanding design. The more specific you get, the more controlled the outcome. So while Midjourney closes the gap on quality as it relates to realism, graphic quality, or stylized approaches, the final production will vary drastically if not given very specific prompts driven by key design words.
We also found that, to the untrained eye, there is little to no discernable difference in final production between architecture students and anyone else who may have access to Midjourney. We found this through surveying over 30 subjects of varying ages and backgrounds to try and identify which images were produced by the design students. Even though some images were picked slightly more than others, the answers were not consistent, and when asked how confident the subject was in their answers on a scale of 1-5, nobody answered above a 3. This goes to show that most of the subjects were taking their best guess and could not tell a difference. Lastly, when considering the complexity of the prompt, it becomes much easier to identify design students. With almost every prompt doubling the length of their untrained counterpart and a vast vocabulary for design and architectural principles, the prompts, when laid out, are easily distinguishable as to who wrote them. This further proves the point that while the final production quality remains consistent across the test of whoever writes the prompts, the design students tend to narrow the window and achieve a specific goal with their tailored prompts.
So those results beg the question: what components create a strong architectural prompt? Here is what we found to be the important factors in crafting a prompt for AI generators such as Midjourney.
2D/3D
Firstly it is important to establish whether you want a 3d rendering or a 2d drawing. This will drastically change the outcome of your prompt.
Medium/Material
After you have chosen 2d or 3d, it is important to include how you want this image to be represented. If 2d, include material like paint, pencil, canvas, paper, realistic image, etc . . .
Architectural Element
This is where you define the subject of the image. This could be a single-family home, a bridge, a painting, or any other subject of the image.
In The Style Of…
This is where you establish a known designer and include “in the style of (designer)”. This is one of the most powerful parts of the prompt that will really pinpoint the style you are aiming to achieve. Finding a designer that has the style you want to emulate is arguably the most impactful portion of the prompt.
Architectural Style
This is similar to the last prompt but refers to a design style rather than a designer themself. This could include a design principle, time period, or architectural style.
Description of Design Elements
This portion of the prompt is where you can get into the details you want included in your design, whether this is windows, structure, colors, or other design elements, this is where to include them.
Example Prompt
Create a photorealistic 3D Rendering of a bridge in the style of Zaha Hadid. This should be a suspension bridge that is primarily white and utilizes unique lighting across a beautiful river. The bridge should be futuristic and kinetic. The season is Fall, the time of day is dusk, make it look dramatic.
Our Research
The research portion of our study was comprised of testing multiple software and how they can be used in the field of design. As AI continues to develop there are more and more that are specifically for design purposes. One of these software that has caught the eyes of many designers is called Prome AI, This software is specifically designed to create renders and drawings from sketches, or vice versa, create sketches from photos or renderings. This growing platform cuts the design visualization and production phase down to a fraction of the usual time needed to create a high-quality rendering. Prome has established itself as a design-driven AI software alongside others such as Maket AI and the Artistic Inspiration Renderer (AIR) for Sketchup. These software and others continue to bleed into other steps of the design phase beyond just visualization. Maket prides itself on the ability to create floor plans from a list of room types and dimensions. AIR for Sketchup is able to take simple blocking diagrams and transform them into beautifully designed buildings. This allows the visualization step to be achieved almost anywhere in the design process.
Prome AI Experiment
Thinking back to our first year in architecture, we wanted to see if this newly developed software would have been beneficial within the design process. Replicating a project done early in our academic careers, we took an abstract painting and extruded the shapes to form a design, using the geometrical painting as guidelines. While this was certainly an abridged version, we were curious to see what AI software could do with simple extruded shapes. We asked AI to create “urban skyscrapers” for our blocking design. Prome AI features a setting in which you can control the creative freedom the software has, meaning it can stick strictly to the geometry you provide or branch out for a more concept-driven piece. Below are the two experiments we conducted. First, the abstract piece of art followed by two views of the blocking diagram and extruded shapes as a result of the painting. Using Prome AI, we imputed these simple screenshots to see the result we were able to achieve using simple line drawings as a base. Attached are the three options for each view that we felt were the strongest ranging from being as precise the geometry as possible to giving the software more creative freedom.
Our Findings
While the renderings produced were certainly not as strong as possible, we were shocked by the quality given there was no prompt provided and or post-production editing. We have continued research with this platform and simple changes could be made that would likely result in stronger results such as specifying a material or site or adding more detail in the initial sketch; however, for this experiment, we wanted to see what the software was able to produce with as little information as possible. We can see this being used in a number of different ways whether that be to generate a concept such as shown in the experiment above and/or produce final renderings by uploading a 2D floor plan or more developed screenshots from your modeling software. We believe this to be one of the stronger AI platforms and expect it to be utilized more frequently in the future given that it is one of the few free AI-based design websites.
How Can Architecture Students Learn Through AI
Those who are familiar with AI understand how quickly and relatively accurate these programs can be when given a prompt. Most students, regardless of their major, use ChatGPT to answer test questions and write their essays. It is a quick and easy software to obtain work that would take a student hours to achieve. Design students are familiar with DALL E and Midjourney. These are AI programs that work just like ChatGPT. A prompt is given that describes the intended image and the AI will produce a quality image in that liking. Most students use these programs to produce work or an idea. However, what if these AI programs were used to better understand the design of architecture? Architecture is a vast field of design with different styles and vocabulary. First-Year students are introduced to architecture and interior design by a series of exercises of hand-drawn work. They are then given some history about architecture and eventually they get to design their first structure. An issue about this is that they don’t get a taste of the vocabulary and styles of architecture that might interest them. A solution to this is to allow and promote the usage of AI programs like ChatGPT to assist in the learning of these architectural styles and terms. We believe an early introduction to AI for students could assist and benefit their design careers. Here is an example of what the combination of ChatGPT and Midjourney could achieve for students:
Exercises like this can allow the student to understand what makes a building follow the brutalist style of architecture. Instead of giving Midjourney a basic prompt like “Create a brutalist pavilion”, the student can learn and practice different architectural style and terms and start to understand different components of a building. Here is an example of how a student can start to understand the spatial layout of a building.
Questions like these might be common sense to those with experience in architecture. However, this could be a concept that first-year architecture students never get the grasp of until later in their education career. Personally, there are things that school didn’t teach me that I later learned by being in an architecture office. They seem like simple things, yet we were never taught them for whatever reason. Allowing a student to learn architecture through AI can boost their learning and prepare them for real work outside of the classroom. It will be interesting to see how AI can influence an eager student who is interested in design.
What Does the Future Hold
As AI becomes more advanced, and especially as it is developed with specific design intent, it will change the industry more and more. As far as we can tell, AI will never replace the designer completely; however, it can aid in the process and significantly speed it up. It is difficult to say how much AI will be adopted into the professional field of work, especially at this early stage where it is changing so quickly. We strongly believe, that at least one person in every office, regardless of size, should be keeping up to date on AI news in design and what changes are happening and coming in the future. This is a tool that is far too powerful and full of potential to go unused. We understand that this topic is uncomfortable for a lot of people who feel their jobs may be at risk down the line as AI expands. We urge you to not think of it as a replacement and use it as a tool. Take this as an example: Say you are in a firm and you specialize in conceptual design and design visualization. Without the use of AI, you are able to produce 3 conceptual designs and 3 renderings in a day. Now with the use of AI, you are able to 10x that amount and create 30 conceptual designs and 30 renderings in the same amount of time. Let’s say that half of the AI-generated or aided designs do not meet the standards you need them to, you have still increased your work by a multiple of 5. As you get better at crafting prompts and using new design-based AI you may be able to get that success rate closer to 100 which means you have increased work by a multiple of 10. You, as a worker who understands how to implement AI into certain parts of your work and manipulate and massage it into a design you envision, have become multiple times more valuable to the company you work for. This technology is unavoidable, so we recommend getting on the right side of it and understanding it from the beginning because no matter how many people say AI is going to take over the world, the much more likely outcome is that it revolutionizes the way we work and the output we can achieve.
2D Images to 3D Model
Our initial exploration delved into the functionalities of various AI software aimed at transforming 2D images into dynamic 3D models. Building upon our prior investigations, we turned to MidJourney, a generative AI platform renowned for its ability to generate images based on text prompts. Using this technology, we created a series of experiments, generating abstract images and seeing how they might translate into tangible 3D models. Beginning with a straightforward prompt such as “realistic image of an abstract sculpture” within MidJourney, we generated the image showcased below as a starting point for our analysis.
Employing this image as our foundation, we fed it into several AI software platforms to see what the output may be. The trio of softwares we used to compare included HuggingFace (https://huggingface.co/spaces/jiawei011/dreamgaussian), 3D SCM (https://3d.csm.ai/detail/SESSION_1708094531_3139943), and Tripo (https://www.tripo3d.ai/). Below images outline our observations and insights gained from each program.
Our exploration revealed Tripo and 3D SCM as standout performers among the three software investigated. While HuggingFace may suffice for smaller and simpler images, Tripo and 3D SCM delivered notably superior results overall. Tripo demonstrated a quicker processing time compared to 3D SCM, yet with slightly more imperfections in its models. Conversely, 3D SCM required more time for model completion but created fully enclosed meshes that appeared to be more complete. Both options exhibit promise, with the choice likely hinging on the specific needs for the final model. It’s worth noting that some level of model refinement may be necessary for both softwares.
When using these platforms for larger, full-scale images such as full buildings, the final output significantly diminished in quality. Thus, it is clear that these tools excel more with small, singular objects as there tends to be confusion when dealing with a full scenic background.
As we continue to explore additional options in this evolving landscape, Tripo is the overall preferred choice for generating clean 3D models from images, easily exportable as OBJ files for seamless integration into your projects. While these tools are still evolving, their potential impact is evident, offering the convenience of placing specific products directly into projects without laborious modeling efforts or swiftly generating complex 3D models to be placed in the background for renderings. Although these software solutions are relatively new, their anticipated significance in design workflows is undeniable.
Text to 3D Model
In addition to transforming 2D images into 3D models, numerous software options also facilitate the conversion of text prompts into 3D representations. While both processes may seem similar, they serve distinct roles within the design workflow. While the image-to-model conversion is typically used for final production, text-to-model conversion proves to be valuable for early-stage problem-solving and schematic design, while also being relevant for final production purposes. Whether you are looking for a conceptual foundation for a new project or just to populate rendered models with objects, text-to-model conversion is a strong choice. Once again, we embarked on an exploration of multiple software platforms to discern which offered the best results. The three softwares investigated were Meshy AI (https://www.meshy.ai/), SudoAI (https://www.sudo.ai/), and Tripo (https://www.tripo3d.ai/). This evaluation involved imputing a simple prompt into each platform to assess the outcomes. The prompt used for this test was “a soft cream-colored couch,” resulting in impressive outputs across the platforms. While variations in style, size, and design were anticipated, our objective was to gauge the general output of each platform. The subsequent section presents the results obtained from all three platforms.
As you can see, each test resulted in highly detailed models. The softwares excelled at interpreting simple prompts, yet encountered difficulties when presented with more detailed object descriptions. We envision this software playing a pivotal role in schematic design, aiding in the determination of spatial compositions and the visualization of specific objects within a given space. Upon inspection, all models from each software exhibited remarkable detailing. We eagerly anticipate the continued evolution of this tool, recognizing its potential to streamline design processes by alleviating the time-consuming task of manual modeling, particularly for intricate furniture pieces or components not readily available in rendering software libraries.
Photogrammetry
Continuing along a similar topic, photogrammetry involves capturing images or scans of real-life objects or buildings to generate 3D models. One notable software facilitating this process is LumaAI. Captivated by its capabilities, we embarked on an in-depth exploration of this software. We began by installing the Luma app, which gives users the ability to scan objects comprehensively, capturing them from every angle and transforming them into intricate 3D meshes. Using a technique known as NeRF, this process ensures precise scanning. Our experimentation involved scanning two items: a stool and a previous architectural model crafted for our studio. Below, you’ll find screenshots depicting the scanning process within the app, alongside the exported meshes in OBJ format imported into Rhino. Additionally, we endeavored to capture a conventional video circling around a water bottle, albeit not through the app, thus lacking real-time guidance on camera movements. The primary aim of this endeavor was to assess whether Luma could be used for drone shots of structures and seamlessly process them. The outcomes were promising; despite minor imperfections that could be refined in post-production, both methods yielded comparable results, underscoring the software’s potential for scanning full-scale architectural structures.
Final Demonstration
Utilizing the tools we researched during our semester-long study, we embarked on a final demonstration to showcase the power of these tools and their seamless integration into the design workflow. Our journey commenced with a simple geometric form—a rectangle—exploring the transformation of this basic shape into a simple café scene using predominantly AI-generated objects. Below, you’ll find a Twin Motion image export illustrating the space.
Every object within this scene, excluding the walls, windows, ceiling, and people, was generated using AI software, whether through text-to-model or image-to-model platforms such as Tripo, Sudo, and Meshy. The results speak volumes: while some fine-tuning in Rhino was required, the majority of objects were directly imported into Revit from AI software. Admittedly, there are areas that warrant refinement, such as material application and polygon smoothing, yet the overall outcome signifies a significant technological leap with profound implications for the design realm. While our focus here primarily revolved around incorporating background elements and entourage into the space, these tools hold immense potential for modeling bespoke specifications or intricate objects that would otherwise demand extensive time and expertise to create. Given we are not interior designers, the space is not nearly complete, just an idea of what these objects will look like within a final presentation. The subsequent video provides a brief walkthrough of the space, offering insights into how these objects can be integrated into the virtual realm of VR and 360-degree views, enriching the design experience.
This demonstration underscores the remarkable precision achievable through these software solutions and shows the ways in which they can be used by designers moving forward. We eagerly anticipate the continued evolution of these tools and remain committed to integrating increasingly innovative AI technologies into our design processes, ensuring continued exploration and advancement in the field.
Where Should I Start
There are plenty of social media pages, blogs, and news outlets that are constantly covering developments in artificial intelligence. We recommend finding one or two that you enjoy and following them to stay up to date on the latest news, this could be 5 minutes a day glancing over new articles.
Next is to experiment with larger, more fine-tuned, and available AI resources such as Chat-GPT, Bard (google), Dalle, SIri, and Midjourney. These more developed versions of AI are easy to use and understand and are largely free to the public.
Lastly, after you have familiarized yourself with some of the capabilities of the more mainstream intelligences, you should start to research your line or work or a niche that interests you and find what AI software are being developed for that specific purpose. In our case of architectural design, we recommend you begin experimenting with some of the more novel software we spoke of such as Prome AI, Maket, or AIR (just to name a few). Don’t just test these to test them, try to actively think of if and how it could be implemented into your workflow or design process.
Conclusion
The inevitability of AI’s integration into every facet of our lives is palpable. Rather than shying away from this transformative technology, we must embrace its potential and actively explore its applications within our respective fields. In the realm of design, we’ve witnessed the emergence of numerous AI platforms—from MidJourney to PromeAI—each offering unprecedented capabilities, ranging from aiding problem-solving in schematic design to facilitating the creation of final construction documents. This rapid advancement presents an array of opportunities to streamline and enhance the design process, compelling designers to adapt and harness the potential of AI rather than retreat from it.
Our journey throughout this year has afforded us invaluable insights and practical experiences, equipping us with the skills and knowledge to navigate this evolving landscape with confidence. As we transition into the professional realm post-graduation, we are poised to bring a fresh perspective informed by our exploration of AI technologies. By embracing innovation and remaining proactive in our engagement with emerging technologies, we are poised to contribute meaningfully to the evolution of our field, ensuring that we remain at the forefront of design innovation and excellence.
Alex Hamilton
University of Cincinnati, DAAP | Bachelor of Architecture | 2024
Jackman Land
University of Cincinnati, DAAP | Bachelor of Architecture | 2024
Nathan Rosenberger
University of Cincinnati. DAAP | Bachelor of Architecture | 2024
More information on AI and Prof. Ming Tang’s research:
Ming Tang has conducted previous AI investigations, including Building Safety Analysis with Machine Learning, AI-driven design automation, and human-machine symbiosis for creative arts.