The cafe sign in the approved render says the wrong name. The letters look almost right, the symbol has the right color, and the client notices in under a second. Another generation fixes the word but changes the canopy. A third preserves the canopy and invents a second door.
This is not a prompting problem. It is a control problem. The workflow asked a generative system to redraw an asset that already existed in exact form.
Chaos added an Insert Images feature to Veras 5.0 that points toward a cleaner method. Its product material says users can place reference photos, logos, and entourage on the 2D canvas, then move, scale, and rotate them. The canvas arrangement is saved, and a user can point or draw at a selected area while describing a local change. Those are vendor-stated capabilities, not results from our own test. The important idea travels beyond one product: exact assets should enter the image as controlled inputs.
Decide what may be generated
An architectural image contains several kinds of visual information. Treating all of them as prompt material gives the model permission to reinterpret facts that should remain fixed.
| Element | Best source | Allowed change |
|---|---|---|
| Atmosphere, weather, mood | Prompt and references | Broad exploration |
| Generic planting or crowds | Prompt, library, or controlled insert | Composition within brief |
| Approved logo or artwork | Exact supplied file | Perspective and lighting only |
| Wayfinding message | Typeset artwork or model geometry | None without approval |
| Specified product | Approved photo, asset, or model | Placement and view |
The boundary is simple. If the object has an owner, identifier, approved spelling, product reference, or contractual meaning, do not ask the model to remember it. Supply it.
A prompt describes a class of thing. An insert identifies this exact thing.
Build an exact-asset packet
Before opening the rendering tool, collect the items that cannot drift. A small packet prevents the team from pulling an old logo from a web search or using a low-resolution screenshot because the approved file was difficult to find.
- Source file: the approved vector, transparent raster, photograph, or rendered product view.
- Identity: client, brand, artwork, product, or sign family name.
- Revision: issue date or version supplied by the owner.
- Usage: where the asset may appear and whether it is cleared for public presentation.
- Placement note: intended wall, plane, crop, orientation, and approximate dimensions.
- Do-not-change note: protected text, proportions, color relationships, and exclusions.
Keep original and working files separate. A flattened PNG prepared for a canvas may be the practical input, but it should link back to the approved source. Name the working file with its revision instead of calling it final-logo.png.
Place first, generate around it
1. Lock the architectural view
Begin from a named camera and a saved model revision. Record the image dimensions and crop. If the camera moves after placement, the insert may no longer align with the surface, and a reviewer can mistake a perspective error for a design change.
2. Establish the plane
Identify the surface that owns the asset: storefront glass, a reception wall, a freestanding sign, a digital display, or a floor graphic. Use corners, joints, or measured dimensions as anchors. Scale against the actual opening or modeled sign zone, not against a generated person.
3. Protect the asset
Mask or isolate the exact graphic before broad generative edits. The tool may need to integrate edge softness, reflections, shadow, or color temperature, but it should not redraw the protected center. After every pass, compare the insert with the supplied source at 100 percent view.
4. Add physical integration deliberately
A pasted logo can be exact and still look false. Decide what makes it part of the scene. A vinyl mark on glass may pick up reflections and partial transparency. Raised letters cast shadows and need plausible depth. A printed mural follows the wall plane and material texture. Integration is a separate operation from identity.
5. Keep a clean composite route
If the generative tool cannot preserve the asset, stop sending it through the model. Produce the environment without the protected graphic and composite the approved source afterward in an image editor. A modest manual step is cheaper than approving a convincing counterfeit.
Apply the method beyond logos
Brand marks are the obvious case because errors are visible. The same rule applies to any exact visual element.
Wayfinding: room numbers, arrows, accessibility symbols, exit messages, and directories carry operational meaning. Generate the setting, not the message.
Artwork: an artist's piece has authorship, crop, color, and licensing conditions. Use the approved reproduction and record its source. Do not describe it loosely and publish a synthetic imitation under the artist's name.
Specified products: a named chair, luminaire, appliance, or facade product may be important to client approval. Insert an approved asset or model view. If the image uses a generic stand-in, label it as such.
People and uniforms: a generic crowd can be generated within a representation brief. A named person, staff uniform, school crest, or safety garment needs closer control and permission.
Maps and diagrams: site maps, leasing plans, evacuation graphics, and data displays should originate in the project record. A generative facsimile can look authoritative while changing labels or boundaries.
Review at three scales
First inspect the full composition. Does the asset sit in the intended place and support the image hierarchy? A perfectly reproduced logo can still dominate a facade or float above the material.
Then inspect at presentation size. This is where spacing, contrast, reflections, and perspective need to work. Check the actual PDF or slide size, not only the editing canvas.
Finally inspect at 100 percent. Compare protected text character by character. Check mark proportions, brand colors, cropping, repeated signs, mirror errors, and accidental duplicates. If the asset carries required wording, have someone other than the image maker read it.
Save the approved render with the source asset revision and the composite or canvas file. If the logo changes next week, the team should be able to replace it without regenerating the entire building.
Our take: generation needs a stopping line
The new Veras canvas feature is notable because it recognizes that references sometimes need coordinates, not just influence. Moving, scaling, and rotating an image is ordinary graphics work, but putting that control beside AI generation is the point. It gives the operator a place to say, “this file, here.”
Studios should make that phrase part of their render brief. Name every element that can vary, every element that must match, and the file that governs each match. Use AI for atmosphere, options, and visual repair. Use controlled assets for identity.
If the client already supplied the logo, the model does not get to redesign it.
Editorial basis: the 26 September 2026 ArchiGen AI intel sweep and Chaos's Veras 5.0 release material, which describes Insert Images for reference photos, logos, and entourage, with move, scale, rotate, saved canvas placement, and targeted markup. Those capabilities are vendor claims. This article reports no hands-on test; the packet, placement sequence, protection rules, and review gates are editorial recommendations.