Every practice has lived through the same awkward week. The client has seen the plan, they have AI Image Models nodded politely at the section, and then they ask the question that no drawing answers: what will it feel like to stand there? Traditionally the honest reply is “give us ten days and a visualisation budget.” Increasingly, it is “give us the afternoon.”

Generative image models have quietly moved from novelty to normal inside design studios, and the interesting part is not that they make pretty pictures. It is where in the timeline they sit. Nobody serious is using them to replace a final photoreal render. They are being used earlier — in the messy stretch between the first massing study and the first proper visualisation, where the cost of producing an image has always been higher than the value of any single image.

The economics of the early image

A traditional render is expensive because it is a commitment. Someone has to build geometry, assign materials, place lights, wait for the farm, and then AI Image Models someone has to explain to the client why changing the brick bond means starting again. That cost structure pushes studios to produce few images, late, and to treat each one as a small deliverable.

Generative models invert that. The marginal cost of an image drops to cents, which means the correct number of images is no longer three, it is three hundred. You are not producing a deliverable, you are producing a search. Show a client thirty variations of the same courtyard elevation with different fenestration rhythms and you learn more in twenty minutes than in three rounds of formal review — not because the images are accurate, but because people are far better at reacting than at specifying.

This is the honest framing to keep in mind: these tools are instruments for finding out what the client actually wants, not for documenting what you have already decided.

Where the current generation got noticeably better

Two limitations used to make these models unusable for design work, and both have eased.

The first was text. Older models produced signage, room labels and dimension callouts as convincing gibberish, which meant any board with legible annotation had to be composited by hand. Current models AI Image Models handle short legible text reliably enough that a wayfinding study or a facade with real shopfront signage survives without retouching.

The second was consistency. Early on, each generation was an independent roll of the dice — you could get a beautiful lobby, but not the same lobby from a second angle, which made the technology useless for anything narrative. Reference-conditioned generation changed that. Feed a model your massing screenshot, your material palette, or a photo of the actual site, and the output stays recognisably tethered to your project rather than drifting into generic Pinterest modernism.

That second capability is what moved these tools from marketing departments into design teams. An image that ignores your project is decoration. An image constrained by your project is a design tool.

The workflow that actually holds up

The studios getting value from this share a fairly consistent pattern, and it looks nothing like typing a paragraph and hoping.

Start from your own geometry. A grey-shaded viewport capture from your modelling package, used as the structural reference, keeps proportion, sightlines and site context honest. Prompting from a blank page produces images that look like architecture but do not describe your building.

Prompt materials and light, not style. “North-facing, overcast, late afternoon, lime render, oiled oak joinery” gives you something you can interrogate. “Award-winning stunning architecture” gives you a mood board you cannot act on.

Generate in families, not singles. Fix everything and vary one parameter — window proportion, ceiling height, floor material — across a dozen outputs. A family of images is an argument; a single image is just a picture.

Label them as studies. The failure mode that damages trust is a client who believes an AI study is a commitment. Watermark early-stage imagery, say plainly that dimensions are indicative, and keep the boundary between exploration and documentation visible.

The integration question

Most practices begin inside a consumer chat interface, which is fine for a partner exploring on a Sunday and useless at studio scale. The moment you want a hundred variations tied to a project folder, named by option number, or generated from a BIM view on a schedule, you need the model behind an API rather than a chat box.

That is a procurement decision more than a technical one, and it is worth understanding the pricing shape before committing, since image models are billed per generation rather than per seat and the cost curve of a three-hundred-image search is very different from a subscription. Studios comparing options typically look at per-image cost, the ability to pass a reference image, and whether the same endpoint can be swapped between models as they improve — the current rates and capabilities for the GPT Image 2.5 API and competing image models are a reasonable place to start calibrating that budget.

The practical advice: do not sign anything annual. This field is moving fast enough that the model you standardise on today will be second-best within two quarters, and the only durable decision is to keep your workflow model-agnostic.

What this does not solve

It is worth being blunt about the limits, because the enthusiasm is running ahead of the evidence.

These models do not understand structure. They will happily produce a cantilever no engineer will sign, a stair with inconsistent risers, and a junction that cannot be built. They have no concept of code compliance, fire strategy, or thermal performance. They produce images that are plausible, not buildable, and the distinction matters enormously once someone downstream treats the output as intent.

They also carry a strong aesthetic prior. Left unconstrained, they drift toward a particular contemporary vocabulary — pale concrete, floor-to-ceiling glazing, an improbable amount of vegetation — because that is what dominates the training data. If your practice has a specific material language, you will spend real effort pulling the model away from the mean, and reference images are the only reliable lever.

And they do not replace the visualiser. The final image that goes in the planning submission or the competition board still benefits from someone who understands how light behaves in that specific room at that specific latitude.

The reasonable position

The studios doing this well have landed somewhere unglamorous: generative imagery is a fast, cheap way to have a better conversation earlier, and a poor way to make a promise. Used in the first third of a project it compresses weeks of misunderstanding into an afternoon. Used in the last third it creates expectations nobody can meet.

That is a smaller claim than the marketing makes. It is also the one that survives contact with an actual client.