Yesterday we wrote about the new engine under Veras 4, the move to the Nano Banana 2 generation (Gemini Flash Image 3.1 Preview), and what cleaner output and higher quotas mean for throughput. That's the part that makes the demos look good. But faster renders were never the thing holding architects back from trusting AI visualization on real client work. The thing holding them back was control, the quiet, maddening way every variation drifts: a window appears where you didn't draw one, a parapet shifts, a clean massing study comes back subtly reinterpreted. Veras 4 ships two features aimed directly at that, and they're the reason to pay attention this week.
Design Lock and Image as Input aren't flashy. Neither generates a hero image on its own. What they do is change the supervision cost of AI rendering, how much you have to watch, correct, and re-run to keep the output faithful to the design you're actually presenting. We tested both over a week of live exterior and interior work.
Design Lock: the building stops moving
Design Lock is the headline feature, and it addresses the oldest complaint about every prompt-driven renderer: the model treats your geometry as a suggestion. You generate a variation you like, run three more to compare, and discover the AI has been quietly redesigning the facade each time. For concept mood boards that's tolerable. For anything a client might mistake for the proposal, it's a liability.
With Design Lock engaged, Veras holds the modeled geometry, the massing, openings, and structural lines, fixed while it restyles everything else. You can push the material palette, the lighting, the time of day, the entourage, and the building underneath stays the building you drew. On our test scenes, the difference was immediate: a row of variations that, for the first time, read as the same project in different conditions rather than four cousins of the same idea.
Preserves modeled geometry, massing, and openings across render variations so the AI restyles the scene instead of redesigning it. The control architects have wanted since AI rendering started drifting facades. Strongest on clear, well-modeled exteriors and interiors.
Where it earns its keep
The payoff isn't a single image, it's the batch. Because the geometry holds, you can finally use variation the way you want to: generate eight options of the same approved massing in different light and material treatments, and present them as a coherent set. We used it most for early client reviews, where the value is showing range without re-litigating the design every frame. Run a sweep, pick the two that land, and you're not auditing each one for a hallucinated window before you dare put it on screen.
For the first time, "make me four variations" doesn't mean "make me four different buildings."
The limits worth knowing
Design Lock is strong, not absolute. On clean, unambiguous geometry it's remarkably faithful. On complex or visually busy scenes, or where the model itself is rough, it can still let small details drift, a mullion pattern that simplifies, a soft edge the AI rounds off. It is a dramatic reduction in drift, not a guarantee of pixel-faithful structure, so a final human pass before client sign-off remains the right habit. Treated as the control it is, rather than an autopilot, it changes how much you can trust a batch you didn't watch generate.
Image as Input: show, don't prompt
The second feature solves a different control problem, the one where you know exactly the material or mood you want but can't get there through prompt language. Image as Input lets you hand Veras 4 a reference photograph and have it extract the materials, textures, and colour atmosphere from that image to drive your render. Instead of typing "warm travertine, late afternoon, soft Mediterranean light" and hoping, you show it the photo and it reads the character off the pixels.
In practice this is the more immediately useful of the two on a day-to-day basis. Prompt language is a lossy way to specify a look, two people read "moody concrete" completely differently, and so does the model on two different runs. A reference image collapses that ambiguity. We fed it everything from a client's mood-board photo to a snapshot of an actual brick we'd specified, and the render picked up the palette and surface character far more faithfully than any wording we tried.
Feeds a reference photograph to Veras and extracts its materials, textures, and colour atmosphere to drive the render. Replaces lossy prompt language for matching a specific look. Best for hitting a client's reference or a specified material exactly.
The reference photo decides the result
As with any image-conditioned tool, the input governs the output. A clean, representative reference, one whose materials and light are actually what you want, produces a render that inherits that character cleanly. A busy, mixed-signal reference (three materials, two light temperatures, a dominant foreground object) gives the model conflicting cues and a muddier result. The discipline that paid off for us was choosing references that isolate the one thing you're after: a flat material sample for material, a clean atmospheric shot for mood, rather than one photo trying to carry both.
The standout use is the client-supplied reference. When a client sends "we want it to feel like this," you can now act on the image directly instead of translating it into prompt text and losing half of it in the process. That's the difference between approximating their reference and matching it.
Two controls, one problem
| Dimension | Design Lock | Image as Input |
|---|---|---|
| What it controls | The geometry stays put | The look matches a reference |
| Problem it solves | Facade drift across variations | Lossy prompt language |
| When you reach for it | Generating a batch to present | Matching a specific material or mood |
| Biggest dependency | Quality of the modeled geometry | Quality of the reference image |
| Trust for client work | High, with a final check | High on clean references |
Read together, the two features describe where Chaos thinks AI rendering still loses to a careful human: not in speed or photorealism, both of which are largely solved, but in fidelity to intent. Design Lock defends the intent of the design; Image as Input defends the intent of the look. Both are about removing the gap between what you meant and what the model returned, and that gap, not render time, is what kept AI rendering parked at "concept only" for most practices.
Our take: control is the feature that actually matters
It's easy to be cynical about another point release with AI features, and we usually are. But there's a real shift here worth naming. Through 2024 and 2025, AI render tools competed on output quality, who could make the prettiest single image. That race is mostly over; everyone's output is good. The 2026 competition, the one that decides what architects actually adopt into a billable workflow, is control: can I get the result I intended, reliably, without babysitting every frame? Design Lock and Image as Input are Veras planting a flag on exactly that question.
If you already run Veras, both features are immediate keepers, because they reduce work you were already doing, re-running drifted variations, fighting prompt language to hit a material. If you've held off on AI rendering because you couldn't trust it near a client, this is the release that should make you re-test your assumption. The engine got faster, fine. The more important news is that the building stops moving.
What we'll be watching is how Design Lock behaves as scenes get genuinely complex, dense urban context, intricate facades, mixed interior-exterior views. The conservative read is that it's a large step, not a finish line. But it's the right step, aimed at the right problem, and that's more than most AI feature drops can claim.
If you run Veras this week
Turn on Design Lock and generate a batch of the same approved massing in four lighting conditions, then check whether the openings held, that's the test that tells you what it's worth on your geometry. Use Image as Input on the next client reference instead of trying to describe it, and compare the result to your best prompt attempt. And keep the final human check before client delivery; these are controls that earn trust, not autopilots that replace it.
We test new AI rendering features on real project work and publish the honest version. Join the studio newsletter for weekly field notes, or read our take on the Nano Banana Pro engine behind Veras 4 for the throughput side of the story.
Tested by Vista Studios on live project renders in Veras 4. No affiliate relationship with Chaos or Veras. Design Lock evaluated on modeled exterior and interior geometry; Image as Input tested with client-supplied references and specified material samples.