A single exterior image can imply a convincing building while saying almost nothing about the wall behind the camera. A floor plan can define rooms while leaving ceiling heights, roof form, structure, and material junctions unresolved. Turning either source into an editable scene therefore requires interpretation, not extraction.
That distinction sits behind the newest claim on the Veras product page. Chaos says GPT-6 Astra can read images, sketches, text, or floor plans, match camera angles across references, model walls, roof, structure, and materials, and place the result in Revit, Rhino, or SketchUp. The same page says Cosmos assets are used where matches exist. It also lists proxy baking for lighter host models and first-person navigation.
Those are vendor claims, not findings from an ArchiGen test. They describe a larger object than the image-to-3D asset workflow Veras introduced earlier. A chair can be accepted or rejected by itself. A scene creates relationships among hundreds of parts. The review must grow with the claim.
A scene can be visually coherent and technically contradictory at the same time.
Start by naming the source certainty
Before generation, mark what each source actually establishes. A measured floor plan may establish room boundaries, openings, and dimensions. A perspective photograph may establish one camera, visible materials, and occlusions. A sketch may establish intent without fixed size. A text prompt establishes a request, not evidence.
Keep a short source ledger with four labels: known, inferred, requested, and unknown. If a roof is absent from every input, its generated pitch is inferred. If the prompt asks for timber screens, their presence is requested. If the plan carries dimensions, those values are known. This ledger prevents a polished model from laundering assumptions into facts.
| Input | Strong evidence | Persistent gap |
|---|---|---|
| Floor plan | Plan relationships and labeled dimensions | Heights, unseen elevations, assemblies |
| Single image | Visible silhouette, camera, surface appearance | Hidden sides, depth, exact scale |
| Multiple images | More surfaces and camera relationships | Occluded construction and true dimensions |
| Text or sketch | Design intention | Most measurable project facts |
Run the second-camera test first
Open the generated result from a view that was not present in the sources. Use a plain shaded or hidden-line display before admiring materials. Orbit behind the building. Cut a section through the stairs. Look down from above with the roof hidden. A model that exists only to support one image will reveal stretched backs, paper-thin walls, intersecting floors, sealed doors, or geometry that ends outside the original frame.
Do not repair defects during this pass. Count them. Separate missing geometry from malformed geometry and speculative geometry. Missing geometry is often quick to model. Malformed geometry can poison edits and exports. Speculative geometry may be acceptable for a concept scene if it is labeled, but it must not masquerade as recovered design information.
Check cameras before checking surfaces
Chaos says the system can match camera angles across references. A camera match can be tested. Overlay a source image and a viewport capture at the same resolution. Compare major corners, horizon, vertical convergence, openings, and occluding edges. If several references exist, repeat the overlay for all of them.
A good match in one frame does not prove that the geometry is right. The system may fit an incorrect shape to one projection. Agreement across separated cameras is stronger evidence. Where overlays conflict, preserve the measured plan or model over the photographic guess, and record which source won.
Audit scale with three anchors
Use at least three known dimensions that are far apart: an overall building length, a door opening, and a floor-to-floor height. One anchor can make a uniformly wrong model appear right. Three anchors expose nonuniform stretching and incorrect depth.
Then inspect repeated elements. Count bays, stair risers, mullions, structural supports, and ceiling modules against the sources. A generated scene can land at the correct overall size while distributing that size through the wrong number of parts.
Coordinates need a separate check. Place the scene in a sandbox file with a documented origin, north direction, and level datum. Confirm that imported or generated geometry does not arrive far from the host origin, rotated to a camera instead of project north, or attached to arbitrary elevations. Copy one verified element into a clean file and link the result back. If that ordinary exchange shifts the scene, fix coordinates before anyone begins design work inside it. A beautiful model at an accidental origin is still an expensive coordination problem.
Separate scene geometry from BIM meaning
Editable does not automatically mean native, scheduled, hosted, parametric, or coordinated. Ask what each object becomes in the host. Is a wall a wall, a generic solid, a mesh, or a proxy? Does a door cut its host? Does a floor know its level? Can a repeated window become one controlled type? Does structure carry any analytical or classification meaning?
For an early study, selectable geometry may be sufficient. For design development, the office may need native elements, predictable sections, manageable families or blocks, and clean exchange. State the required status before review. Otherwise every object receives the most generous interpretation available.
Measure the maintenance burden
Record file size, object count, material count, texture storage, and viewport performance before the scene enters a live project. Duplicate materials and high-detail objects can make a generated result expensive to keep. The proxy-bake option listed by Chaos may reduce host-model weight, but it also creates a dependency between a light representation and fuller geometry. Test what happens when files move, collaborators open the project, and the scene is exported.
Make one ordinary change: raise a floor, widen an opening, replace a material, and move a repeated object. Time the work and note how many related parts require manual repair. An editable scene earns its name when edits propagate predictably, not merely when vertices can move.
Use a two-gate acceptance decision
Gate one: concept scene
Require correct overall scale, cameras that match the sources, complete visible geometry, acceptable navigation, no severe model-health faults, and clear labels on inferred content. The scene may contain generic solids, proxies, and generated materials. It must remain isolated from schedules and technical deliverables.
Gate two: project model
Require host-native behavior where needed, approved coordinates and levels, rational object types, controlled materials, acceptable file weight, export checks, and named ownership. Replace inferred components that affect compliance, fabrication, coordination, or quantities. Archive the original generation separately so later edits do not erase provenance.
This is not a demand that every generated concept become construction-ready. It is a way to stop the wrong status from traveling downstream. A fast scene can be valuable precisely because the office does not spend hours promoting every surface.
Our take: the new unit of review is the relationship
Image-to-3D object generation asks whether one thing has the right scale, form, and position. Full-scene generation asks whether spaces, levels, cameras, materials, and objects agree with one another. That is a harder problem and a more useful one.
Start with a copy of a small project, two or more source views, and three known dimensions. Save the untouched output. Run the second-camera, overlay, scale, host-behavior, and edit tests. Publish the defects beside the attractive frame.
The first view sells the scene. The second view tells the truth.
Editorial basis: the 2 October 2026 ArchiGen AI intel sweep and the current Chaos Veras What's New page, checked 2 October 2026. The feature description and host claims come from Chaos. This article reports no hands-on test. The acceptance protocol is an editorial recommendation.