The useful workflow is not “type a prompt and invent a building.” Architecture teams are getting more control by starting from a hand sketch, massing study, Revit model, SketchUp view, Enscape output, or established render and using AI to test a defined visual question.
Start with your actual design. Tell the tool what may change and what must not. Compare several results, then have the design team decide what still represents the project.
Start with the design you actually have
A source model or view gives the system spatial information that a text prompt does not contain. Current AEC visualization tools can work from Revit, SketchUp, Rhino, Archicad, Enscape, drawings, and sketches. The amount of geometric control varies, so teams should test the exact tool and settings against their project stage.
Keep the source view, model version, reference images, prompt or settings, output, and reviewer together. That makes the iteration reproducible and helps the team explain which part came from the project and which part came from the model.
Say what can change—and what cannot
Massing, openings, primary geometry, camera, circulation, or another piece of approved design intent.
Materials, light, landscape, entourage, atmosphere, season, color, or presentation style.
Run several controlled directions from the same source instead of treating one image as the answer.
The project team selects, corrects, and labels the image for its intended use.
Give the tool less freedom as the design becomes real
Early concept work can tolerate more interpretation. As decisions become approved and the image moves toward client communication or a project deliverable, the workflow should preserve geometry more strictly and reduce uncontrolled variation.
Use explicit review questions: Did massing change? Did a window move? Did the image imply a material or assembly the project has not selected? Does the output communicate the current design, or merely a compelling adjacent idea?
Save the setup when it works
A firm does not need one universal prompt. It needs a small set of tested recipes tied to recognizable inputs and uses:
- Sketch or massing study to atmosphere test
- Model view to material and façade comparison
- Existing render to landscape, entourage, or lighting enhancement
- Approved image to controlled presentation variation
For each recipe, record the source requirements, reference library, important settings, variables the user may change, known failure patterns, review checklist, and accepted destination.
Protect the project and the client
Before uploading a model, view, or image, confirm the account, vendor terms, model-training controls, retention, permissions, client requirements, competition rules, and confidentiality obligations. A beautiful output does not resolve a data-boundary problem.
A convincing image is not enough. The team should be able to say what came from the project, what AI changed, and why the image still represents the design.