AI Powered Design Exploration: From Ideation to Realization

I have worked in the architectural design industry for about twelve years, and the part of the

work that has always interested me most is how new forms are found. Over time I reached a

conclusion of my own: new form rarely comes from a single moment of inspiration. It usually

starts from historical references, and new possibilities come out of transforming and varying

them. Seen this way, the early design process is a fairly logical method, not magic.

If the process is logical, it should be possible to describe it as an algorithm. My long-term goal is

a design methodology that moves from the first idea to a finished project with something close

to automation. That is still a distant goal. This workshop, AI Powered Design Exploration: From

Ideation to Realization, tests one part of it: can AI actually help at the very beginning of a

project, when nothing is decided yet, or does it only become useful once a scheme already

exists?

In the offices where I have worked, 3D computation has become steadily more advanced. To

reduce manual labor and build more efficient workflows, we moved away from typing commands

into 3D software one by one and toward Grasshopper definitions and Rhino scripts. These tools

worked well, but almost always after the early design direction had been set. They were good at

controlling geometry once an idea existed. They did not help much with finding the idea.

The starting point of a distinctive form remained an intuitive, human task, and that is where I

kept running into limits. Too often the early options felt familiar, and it was frustrating to see the

same kinds of answers come back.

Around 2023, generative image tools such as Midjourney began to appear. When I started using

them, I saw that they could be useful in exactly that early ideation stage. Midjourney is not a

precise drawing tool, and it is not a design tool, but it is a strong mutation tool. Given a careful

description of a building, it returns variations I would not have drawn myself.

With the aim of building my own database of forms, I started posting daily work on social media.

That practice led, fortunately, to opportunities to write about generative architecture, including an

essay in Domus 1106, guest-edited by Bjarke Ingels. This workshop is the hands-on version of

that work

.Day 1: Midjourney, for searching

The workshop is split in two, because I believe the two uses of AI in architecture require

opposite mindsets. On the first day, we are searching.

Chapter A unfolds in three stages. The first is Precedent and Inspiration . We closely read a

few named buildings and objects, from architecture and from outside it. Rather than describing

them by style, which tends to produce generic images, we annotate them as architectural

anatomy: a terrace, a recess, a lintel, dual frontality, the way the load comes down, a bulkhead,

the size of an opening, the mass of a material. The output is a written anatomy list, not a mood

board. The hard part of working with references was never finding them. It is turning them into

something you can actually work with.

The second stage is Mutation and Variation . The anatomy list is written directly into

Midjourney prompts, sometimes from a single source, sometimes combining two or three.

Traditionally this kind of testing happens in physical models or 3D software, and it is slow. Here,

the goal is a wide field of massing, facade, and spatial options, many of them unexpected. We

are not looking for a finished image.

The third stage is Critical Selection , and this is where you are the designer. The options are

judged against the project brief. You keep one frame and name what survived in it: the depth of

a facade, a roofline, a recess, a structural opening, or simply the atmosphere. If nothing holds,

you return to the precedent or to the prompt. Day 1 closes as a complete ideation study: one

selected image, and a clear argument for why it is the best option.

Day 2: Nano Banana Pro and Kling, for showing

There is a second, unrelated problem that appears later in every project. Even after the design

direction is settled, visualization consumes a disproportionate amount of time. On the second

day, the rule is the opposite of Day 1: we make no formal variation at all.

Chapter B is a separate study; we do not continue the Midjourney image. Participants begin with

a Rhino screenshot: a perspective, an axon, or an interior. A finished render is not required, only

a clear capture. Using Google's Nano Banana Pro, image generation that respects a given 3D

view, we turn that capture into an accurate visualization. We work with a strict hold list: massing,

openings, and proportions stay exactly as they are in the model. Only material, light, camera,

people, and context are allowed to change. Much of the session is about writing edit instructions

that change atmosphere or occupancy without touching the structure, and about testing light,

season, and camera shifts.

The approved still then goes into Kling and becomes a short clip: an approach, an orbit, a

fragment of an interior walk, or a change of light. The same rule applies: do not generate a

different building. Motion simply lets us review the same design in time, without setting up a full

animation pipeline. Day 2 ends with one accurate still and one short animation, both grounded

in the Rhino view.Understanding why Midjourney and Nano Banana Pro behave so differently, one generous and

unpredictable, the other precise and obedient, is itself one of the central lessons of the

workshop.

What to expect

Participants will leave with a clearer picture of their own ideation process, a method for

annotating a building so it can be written into a prompt, and criteria for selecting among outputs

that they can state out. They will also leave with a practical workflow for turning a preliminary 3D

view into a convincing still and a short piece of motion.

A few honest notes before we begin. AI is not a magic tool that solves all design for you. It

saves real time in ideation and in visualization, but what comes out depends on what goes in.

Choosing the right inputs, recognizing a good output, and refining it into a real project are still

human abilities. Material, site, and construction also remain your responsibility; these tools do

not know where the building is or how it gets built.

This is why I think the qualities of an architect have not changed so much as sharpened. An

architect needs a clear sense of the aesthetic they are pursuing, and the eye to judge against it.

With that in place, the right attitude is an active one: to pursue as much efficiency as possible

with new technology, rather than waiting or resisting.

The number of images you generate matters very little. Selection matters. AI is not a

replacement for the designer, and not a shortcut around design. At its best, it is a good partner.