The gap this fills
The costly part of concept design is the interval between a sketch and something a client can react to. A render needs a model, a model needs decisions that have not been made yet, and so the decisions get made against a drawing that only the architect can read.
Generated exploration fills that interval. The same massing at three times of day. The same façade in four materials. The same room with the light coming from the other side. Not to replace the render — to make the conversation that precedes it possible a week earlier.
The second thing it changes is how many options survive to the client. When a study costs an afternoon, you present the one you committed to. When it costs minutes, you can present three and let the client discover which one they want, which is a materially better conversation.
Variants that stay related
One scheme, many studies
The base view is a node. Material, time of day and weather hang off it, so a change to the scheme updates every study rather than orphaning them in a folder.
Comparable by construction
Studies that share a parent share the framing. That is what makes two options genuinely comparable instead of two different pictures that happen to be of the same building.
Traceable
Each image keeps the inputs that produced it. "Which option was this, and what did we change?" is answered by the graph rather than by a filename convention nobody followed.
Reusable across projects
A material and light setup that worked becomes a template. The second competition entry starts from structure rather than from nothing.
What it must not be used for
Use it up to the point where the scheme is decided, then model it properly. The handoff is not a compromise — it is the intended path, and the tools on the other side of it are better at what they do than any generative pass will be.
There is also a disclosure question. If a generated study is shown to a client or a planning committee, say what it is. The industry norm on this is still forming, and being early to the honest version of it is cheaper than being corrected later.
Questions from practices
Can I feed it my SketchUp or Rhino model?
You can use a viewport render or a screenshot as a reference image, which is what most people do and it works well for holding massing and framing. There is no direct model import.
Will it keep my proportions?
Approximately, from a reference view — well enough for material and light studies, not well enough to measure. If proportion is the thing being decided, decide it in the model.
Is it useful for interiors?
Often more than for exteriors, because interior decisions are dominated by material, light and mood — exactly what this is good at — and less by measurable geometry.
Can we present these in a competition?
Check the competition rules, which increasingly address generated imagery directly. Where it is permitted, label it.
Does it understand daylight accurately?
No. It produces plausible light, not simulated light. For any daylighting analysis that informs a real decision, use a tool that actually computes it.