AI ideation inside a visualization ecosystem. Check whether the output you need is a concept image, an asset or editable scene geometry.
Available · vendor documentedDocumentation review · not tested
Where it fits
Explore material direction and spatial atmosphere, then assess whether a selected visual idea can become useful model content.
What goes in
The checked vendor page describes image, sketch, text and floor-plan references.
What comes out
Vendor-described image outputs and native editable 3D scenes for Revit, Rhino or SketchUp, including Cosmos assets where matched.
How to read the capability
A convincing image and a usable model answer different questions. For an early presentation, compare composition and material intent. For model adoption, inspect scale, object structure and how easily the scene can be revised. Keep the original reference beside both outputs.
Limits to keep in view
The search snapshot and live update page differed during this review. GAD therefore does not assign a version-specific launch date or endorse performance claims. Geometry generated from a reference needs design and technical verification.
A practical evaluation checklist
01
Choose a view with unmistakable bay spacing and openings; compare their locations before accepting an image.
02
For a generated scene, measure three known dimensions and inspect whether repeated objects remain editable.
03
Change one material without changing the brief; record unwanted geometry and camera changes.
Read the output, not just the promise. These profiles summarize official documentation, not hands-on product reviews. Evaluation checklists are GAD editorial recommendations. Availability is scoped to the named feature; an experimental feature does not make the entire platform experimental. No paid placement determines this selection.
Veras 5.2: an acceptance checklist for AI-generated model geometry
What to inspect before adopting Veras-generated geometry: scale, native element structure, relationships, assumptions and the next ordinary edit.
The vendor announcement dated 29 September 2026 reports Veras 5.2. This is a documentation-based account, not a hands-on performance test. Native output is limited to the three named host applications in that announcement. The Chinese counterpart is an abridged summary.
A generated architectural scene becomes more consequential when it can enter the working model. The review must then ask how the objects behave, which dimensions are supported and what information has been inferred from the references.
Chaos's dated 29 September 2026 announcement describes Veras 5.2 Agent mode and native baking into Revit, Rhino and SketchUp. The vendor identifies different outputs: native building elements in Revit, editable groups in SketchUp and editable NURBS geometry in Rhino. It also describes question-mark indicators for assumptions made when references are insufficient. These are vendor-documented capabilities; GAD has not tested their reliability. Official Veras 5.2 announcement
Start with a bounded adoption question
GAD recommends a pilot based on a familiar, non-confidential example for which the team has dimensions and design intent. Choose a limited target, such as an interior arrangement or a small pavilion study. Write down what successful adoption would save and what still requires authorship. A model that needs substantial reconstruction may still be useful for ideation, but that is a different outcome from a time-saving production asset.
Keep references with their provenance. Distinguish a measured drawing from a photograph, sketch or generated image. When reference views disagree, resolve the conflict before the pilot or mark it as an explicit unresolved assumption. Extra pictures do not establish an authoritative dimension by themselves.
Check scale and object structure
Bring the result into a separate project copy or controlled test file. Measure a horizontal span, a vertical dimension and a repeated interval against known information. Inspect the origin, orientation and levels used. Record the actual tolerance required for the intended use rather than adopting a universal pass value.
Next, inspect representative objects in the host application's normal tools. Can the team select, identify and edit the intended component? Does its classification support the planned schedule or coordination task? Are repeated elements organized in a way another modeller can maintain? Editable geometry is only part of that assessment.
Test relationships and an ordinary revision
Use a simple revision that the real project is likely to need. Move a window, change a wall height or replace a finish, then examine the surrounding model. Look for detached openings, duplicate surfaces, unintended changes and inconsistent quantities. Check any information needed downstream instead of assuming visual agreement implies data completeness.
Review every exposed assumption and maintain a second list of uncertainties found by the human reviewer. An absence of visible warnings should never be treated as proof that all hidden dimensions or construction decisions were known.
Decide what enters the project
Classify the output as accepted concept geometry, geometry requiring repair or a visual reference only. Save the tested host version, input set, result, corrections and reviewer. Record the time through the first successful revision, including inspection and cleanup. That gives the team a meaningful measure of usefulness and a clear boundary around the work still owed by the designer.
