AI Select: Prompt-Based Render Editing Grows Up in 2026

For years the deal with AI renders was blunt: you typed a prompt, the model handed you a whole new image, and if one thing was wrong — a sofa you liked, a window that finally sat right — you either kept the mistakes or rolled the dice again. Regenerate, lose the good parts, repeat. That loop is dying. The 2026 tools let you point at one chair, one wall, one patch of floor, and change only that. The rest of the scene holds still.

This shift has a name now: region-select, prompt-based editing. Chaos ships it in Veras 4.5 as “Smart Selection” plus Modify and Replace modes. Graphisoft bakes a lighter version into the Archicad AI Visualizer. The idea is the same across both — surgical edits instead of a full re-roll — and it changes how a visualization actually gets made.

What “AI Select” actually means

Old-school AI editing meant painting a mask by hand. You’d grab a brush, scrub over the region you wanted changed, feather the edges, and hope the boundary didn’t smear. Tedious on a clean product shot. Miserable on a cluttered interior with forty objects.

Smart Selection throws the brush out. The tool reads the image, works out what’s in it, and lets you click once to grab an object, a material, or a whole category. Click a wood floor and it selects the wood floor. Hold Alt and it grabs every instance of that material in the frame — all the oak, everywhere, in one go. Under the hood Veras generates selection maps in parallel, caches them, and restores them when you switch between renders, so the second click on the same scene is instant.

You can also describe what you want in plain words. Type “the pendant lights over the island” and the tool builds a selection for exactly that, saves the prompt with a thumbnail, and lets you reuse it on the next render. It’s the difference between masking and pointing. Pointing wins.

Modify versus Replace: two different jobs

Once you’ve selected something, Veras 4.5 asks a sharp question: do you want to change this thing, or swap it for a new idea? Those are not the same operation, and the tool finally treats them differently.

Modify keeps the underlying geometry and edits within it. Change the dimensions, the material, the finish — the chair stays a chair in the same spot, now upholstered in bouclé instead of leather. This is the mode you want for client tweaks, where the design is settled and someone just wants to see it in a different fabric. It respects what’s already there.

Replace does the opposite. It discards the selection and invents something fresh for that patch of the scene. Useful when you’re still exploring — rip out the whole seating area, let the model propose three different arrangements, keep the one that clicks. One mode refines a decision. The other reopens it. Knowing which you need before you click saves a lot of wasted generations.

Where this lives: inside the tools you already use

The interesting part isn’t the feature. It’s the distribution. Veras runs Smart Selection across every host it plugs into — Enscape, Revit, Rhino, SketchUp, Vectorworks, Archicad, Forma, and the web app. So a Revit user and a SketchUp user get the same one-click selection behavior without leaving their modeling environment. No round-trip to a separate AI app, no export-import dance.

The rendering engine matters here too. Veras 4.5’s selection modes are tuned for its Nano Banana engine — Google’s Gemini image model — which handles the “understand the scene, edit this region, leave the rest” job better than the older diffusion setups that treated every generation as a blank slate. Different engine, different behavior. The tool now exposes that choice instead of hiding it.

Graphisoft comes at it from the BIM side. The Archicad AI Visualizer, powered by Stable Diffusion, sits right in the design environment: build a rough concept model, type “a modern office with wood surfaces,” and get a refined visualization off the massing — no detailed model required for each iteration. You need an Archicad license and a decent NVIDIA GPU or Apple Silicon chip, and the source images stay on your machine, which keeps the IP question simple. It’s less granular than Veras on selection, but it proves the same point: the AI now lives where the model lives.

The workflow, step by step

Here’s how a real edit runs in 2026. Start with a render or a screenshot from your viewport — a still from Enscape, say, or a concept frame out of Archicad. Bring it into the AI panel. That’s your source image, and it’s carrying the geometry and camera you already committed to.

Now select. Click the object or material, or describe it. Pick Modify to adjust it or Replace to reimagine it. Write a short prompt for the change — “matte black frame,” “warmer oak, wider planks,” “swap the artwork for a large abstract canvas.” Generate. Because only the selected region is in play, the walls don’t move, the lighting holds, the composition survives. Compare the result against the source, keep it or reject it, and move to the next element.

The small quality-of-life wins add up. Rotate Image lets you reorient a source in 3D — set pitch, yaw, and roll — and generate a fresh viewpoint without touching the 3D model, which is genuinely strange the first time you see it work. Remove Background spits out a transparent PNG for compositing. Describe Source Image, now running on the Florence 2 model, writes a caption of what’s in the frame so you can reuse it as a starting prompt. None of these are headline features. All of them shave minutes off every pass.

Why geometry-locked editing beats a re-roll

The technical reason region editing works is that these tools condition the generation on the source. They’re not dreaming an image from noise. They read the depth, the edges, the layout of what you fed in, and constrain the output to fit — the ControlNet-style approach that quietly became the default. When you edit a selection, the model respects the geometry outside it and often inside it too.

That’s what makes the output usable in front of a client. A full re-roll gives you a pretty picture that doesn’t match the plan — the window’s in the wrong place, the ceiling height drifted, the room grew a column. Locked-geometry editing gives you the actual room, restyled. For anyone who has to defend a render against a set of construction drawings, that gap is the whole game. Pretty is cheap. Consistent is the thing worth paying for.

Where it still falls down

This isn’t solved. Selection is smart, not psychic. On a busy scene it will occasionally grab the reflection of a chair along with the chair, or miss a slice of material that’s in shadow, and you’re back to nudging the selection by hand. The category select — “grab all the wood” — is great until the tool decides a wood-look laminate and a real oak beam are the same thing, which, visually, sometimes they are and sometimes they very much are not.

Replace mode has the opposite failure. Give it too much freedom and it’ll hand you something confident and wrong — a beautiful sofa scaled for a giant, or a light fixture that ignores the ceiling it’s hanging from. The fix is boring but real: smaller selections, tighter prompts, more passes. The people getting clean results aren’t typing one magic sentence. They’re doing five quick surgical edits and stacking them.

And the BIM-side tools like the Archicad Visualizer trade granularity for integration. You get speed and a tight loop with your model, but not the click-one-material precision Veras offers. Pick the tool for the job: concept exploration off a massing model, or detailed restyling of a near-final render. They’re not the same task.

Questions designers keep asking

Do I still need a 3D model, or can I edit a photo?

Both work. If you have a model, render a frame and edit that — the geometry conditioning is stronger and the results stay truer to the design. But these tools also accept a plain uploaded image, so you can edit a photo of an existing space or a reference shot. The Rotate Image feature even pulls new viewpoints from a single flat source with no model at all.

Will the edit change parts of the image I didn’t select?

Mostly no, and that’s the point. Modify mode in particular is built to preserve everything outside the selection and the geometry inside it. You’ll still get the occasional bleed on tricky edges — reflections, thin objects, shadow lines — so always compare against the source before you send it out. Treat the first result as a draft, not a final.

Which tool should I start with?

If you already render in Enscape, Revit, Rhino, SketchUp, Vectorworks, or Archicad, Veras is the low-friction path — it plugs into the host you’re in and the selection tooling is the most mature. If you live in Archicad and mainly want fast concept visuals off a massing model, the built-in AI Visualizer gets you there without a second subscription. Try the one closest to your existing workflow first.

Is my design data safe when I use these features?

It depends on where the model runs. The Archicad AI Visualizer processes locally and keeps source images on your machine, which is the cleanest answer for IP-sensitive work. Cloud-based engines send your image to a server to generate, so read the terms and check whether your renders are retained or used for training. For confidential projects, favor local generation or a vendor with a clear no-training policy.

How many edits does a client-ready render really take?

More than the demos suggest, fewer than the old way. Plan on a handful of targeted passes — floor, then key furniture, then lighting, then the small stuff — rather than one perfect prompt. Each pass is fast because you’re only regenerating a region. The total is still far less time than masking by hand or re-rolling the whole frame until it happens to line up.

Want renders that stay true to your design while you explore the look? See how Pixintellect puts AI-assisted visualization to work for your projects.

Images: AI-Designed