A rendered lobby contains exactly the chair the team wants. It has the right height, soft corners, timber shell, and low upholstered seat. Until now, that chair remained a picture. Someone still had to search a library, approximate it with a generic family, or rebuild it from scratch.
Veras 5.0 changes that handoff. Chaos says its new Image to 3D beta can inspect an AI-generated image or other reference, match the pictured object to an asset in the curated Cosmos library, or generate new geometry when no match exists. The result is placed in Revit, Rhino, or SketchUp at the intended scale and position. The beta is currently listed as free to use without credits.
That is the most consequential item in today's Veras-heavy sweep, even though the search result still leads with video. Image generation made BIM into an input for a picture. Image to 3D makes the picture a possible input for BIM. The direction of travel has reversed.
What Veras 5 actually adds
The release has more than one feature. Chaos lists Grok Imagine engines for images and video, a canvas tool for inserting and positioning reference images, AI-controlled depth of field, rebuilt Rhino synchronization, a 90-day trash and recovery window, three 3D viewer display styles, final-frame reuse for video, and an interface that works in narrow windows.
Those are useful production changes. The Image to 3D beta is different because it changes the status of a generated idea. A facade light, planter, tree, table, or chair does not have to remain flattened into pixels. It can become an object the team can select, move, count, and show from another camera.
Chaos describes three actions: match, generate, and place. Match searches Cosmos for a close asset. Generate creates geometry when a library object does not fit. Place returns the result to the authoring model at the scale and location inferred from the image and model context. Each action solves a real gap, and each introduces a different check.
| Action | What it saves | What the architect checks |
|---|---|---|
| Match | Manual catalog search | Identity, source, licence, dimensions |
| Generate | Early manual modeling | Topology, silhouette, materials, intended use |
| Place | Scale and position setup | Units, host level, rotation, clearances |
A mesh is not yet a building object
The release page calls the returned result real, scaled 3D geometry. That wording matters. It does not promise a native Revit family with type parameters, OmniClass data, manufacturer identity, fire rating, cost code, maintenance zone, or an approved specification. It promises geometry.
For concept work, geometry may be enough. A massing model needs scale, silhouette, and placement long before it needs a product code. A client workshop may only need to test whether a long bench blocks circulation. A visualizer may need the chair to remain consistent across six views. In those cases, bringing the pictured object back into the model removes tedious reconstruction.
Later phases ask different questions. Can the object schedule? Does it cut correctly in plan? Can the team change width without editing vertices? Does it carry enough information for coordination? Is its material a named project material or a stack of imported bitmaps? If those answers matter, the generated object is reference geometry until someone promotes it.
The return trip from image to model is complete when the object can be governed, not when it can be selected.
The six-part acceptance test
1. Verify dimensions against a known element
Measure overall width, height, and depth immediately. Compare them with the source brief and one fixed model element, such as a 900 mm door. Do not trust a visually plausible chair beside a generated table, because both could share the same scale error. In an imperial project, confirm that unit conversion did not turn millimeters into inches or feet.
2. Inspect placement in plan, section, and elevation
A good camera view can hide a floating foot, buried planter, or chair rotated two degrees from the room grid. Open orthographic views. Check the base elevation, host level, insertion point, rotation, and distance from adjacent objects. For an exterior object, confirm the correct survey or project coordinate context before copying it elsewhere.
3. Read the silhouette at three distances
Inspect the object close up, at room scale, and at whole-building scale. Generated geometry may spend thousands of polygons on a cushion seam that disappears in every deliverable. It may also simplify the one arm profile that made the reference worth choosing. Keep what the project can see and remove what it cannot.
4. Test geometry health
Look for open edges, inverted normals, self-intersections, isolated fragments, and microscopic faces. In Rhino, run the usual object checks. In SketchUp, check nested groups, hidden edges, reversed faces, and component behavior. In Revit, determine whether the imported form behaves in views, exports, and worksharing. A clean screenshot is not evidence of a clean model.
5. Reassign materials deliberately
Map the object to named office materials. Reduce duplicate textures, fix physical scale, and remove files that cannot travel with the project. If the source image only implies a finish, label the material as an intent, such as warm oak appearance, rather than inventing a product specification. Visual certainty must not become technical certainty by accident.
6. Decide its status and expiry
Add a parameter, layer, tag, or naming prefix that says GENERATED-CONCEPT or COSMOS-MATCH. Record the reference image and date. Give the object an expiry point, such as end of schematic design, unless the project lead approves it for continued use. This prevents attractive placeholder geometry from reaching a quantity schedule or construction issue unnoticed.
Where the feature earns its place
Image to 3D is strongest when the image contains an object that is important to the design conversation but expensive to reconstruct before direction is approved. Furniture families, decorative lights, planting proxies, retail fixtures, and loose accessories fit that description. The team can test spacing, views, and composition while the decision is still cheap.
It is less suitable for objects whose geometry carries safety, fabrication, or regulatory meaning. Stairs, guards, doors, facade anchors, structural members, and accessible fixtures should begin with controlled geometry and known constraints. An AI-derived version may support ideation, but it should not become the technical source.
The match route is generally easier to govern than the generation route. A Cosmos asset has a library identity and a known origin. A generated mesh needs closer inspection and a clearer status. That does not make one good and the other bad. It changes the cost of acceptance.
Our take: this is a promotion workflow
Veras 5 does not erase the boundary between a render and a model. It gives teams a bridge across it. The useful studio question is not whether AI can make 3D. It plainly can. The question is what evidence an object must carry before the office treats it as part of the project.
Set two gates. At the concept gate, require correct scale, placement, silhouette, model health, and a generated-status label. At the documentation gate, require native behavior, approved materials, required metadata, and a named owner. Most returned objects should pass the first gate quickly. Only a few need to pass the second.
Veras has found a way to bring the chair back from the picture. The architect still decides whether it gets a seat in the model.
Written from the 9 September 2026 intel sweep and Chaos product information available on that date. Veras 5 Image to 3D is described by Chaos as a beta. Features, credit terms, host support, and output behavior can change. Test with a copy of a project before production use.