Your iPhone Can Scan the Bathroom. GPT-6 Astra Can Build It in Blender. Your Contractor Still Needs a Tape Measure.

A new kind of remodel mockup is showing up on X. Someone scans a room with an iPhone, hands the scan to an AI model, and gets back a Blender scene they can rearrange. Want to see the double vanity against the long wall, or the tub swapped for a walk-in shower? Ask for it.

The best-known demo comes from OpenAI. In its architectural visualization post, GPT-6 Astra, running in Codex, builds a house called "Solace" by writing Blender Python scripts, running them, checking the result with Computer Use, and keeping backups along the way. That house is made up, though. The more interesting posts start from real rooms.

Andrew Dunn (@DunnWithAI) scanned his house with magicplan's LiDAR, had Astra pull in the magicplan data and model it, then worked through remodel ideas on that model. Mihovil Ilakovac (@infomiho) took a Polycam scan of his future kitchen and used Claude with Blender MCP to get a measured technical drawing and a render. Joaquin Bonifacino (@jabr7_isaf) gave Blender MCP two house plans and the prompt "Make this house in blender and its interior," and says he had it in about 45 minutes.

These posts are self-reported and don't have much engagement yet. Most of the Blender-AI posts on X are still fantasy houses. Still, the workflow behind the remodel posts is consistent, and none of it is exotic.

How it works, plainly

There are three pieces.

The scan. iPhone Pro models have a LiDAR sensor, and apps built on Apple's RoomPlan (Polycam, magicplan, and others) turn a walk around the room into walls, openings, and rough fixture boxes. You can export a USDZ or OBJ mesh and a RoomPlan JSON file. The JSON is the useful part, because it stores the dimensions as numbers instead of a lumpy mesh (this explainer covers the formats).

The modeler. Blender is free. The AI doesn't drag a mouse around in it. It writes Python that Blender runs. OpenAI's approach skips plugins and has Codex run scripts with blender --python. The other route is MCP, a standard way to let an AI drive an app. Blender now has an official MCP Server (Blender 5.1+), and most X demos use the community blender-mcp bridge.

The loop. You describe what you want, the AI builds it, renders a view, looks at the view, and fixes things. The open-source LiteReality-Agent shows the pattern clearly. It turns a RoomPlan scan into an empty room, then an agent edits that room, renders it, and compares the render against the original capture.

The catches

This is where the tape measure comes in.

It's a mockup, not a drawing. Martyn Day at AEC Magazine found the first models look impressive, but he says there's "no dependable judgement about circulation, clearances or construction." The AI doesn't know your door swing hits the toilet unless you tell it to check.

Changes are harder than the first build. Day's other finding is that the models tend to fall apart when you ask for revisions. Getting the first model is easy. Getting the fifth version without wrecking the first four is the hard part, which is why checkpoints matter below.

Scans are close, not exact. OpenKitchen3D calls a RoomPlan scan "a starting point, not fabrication truth," and says to confirm every wall against a tape measure. A scan also can't see inside the walls. Plumbing, wiring, and framing are unknown.

Materials look right but aren't. A studio review at ArchvizDaily found the rough layout is the strong part ("six hours compresses to well under one"). Materials and UVs, though, are "plausible, not correct," and "Clay stage is not a deliverable." That tile might not be your tile.

The AI's code runs unguarded. Blender's own MCP page warns that it runs the model's code without safeguards and recommends using a virtual machine. Don't point it at a folder holding anything you care about.

It costs time and money. You need a paid ChatGPT plan or API access (Astra's API price is $10 per million input tokens and $50 per million output). The demos above ran anywhere from about 45 minutes to several hours.

Try it Monday

Here's the recipe pulled together from the sources above, including Vagon's step-by-step Astra setup.

1. Scan the room. Use a RoomPlan-based app such as Polycam or magicplan on an iPhone Pro. Walk slowly and get every corner. Export the USDZ (or OBJ) plus the RoomPlan JSON.
2. Check the scan with a tape measure. Measure each wall, the ceiling height, and the door and window openings. Write the real numbers down and fix any the scan got wrong.
3. Set up a sandbox. Install Blender 5.x. Use Astra through the Codex desktop app running scripts, or add Blender MCP. Keep everything in a dedicated project folder, or a VM if you can.
4. Start with an empty room. Give the AI the scan, your measured dimensions, a few photos, and a short written brief. Ask for walls, floor, ceiling, door, and window only. Check it against your numbers before going further.
5. Add fixtures in passes. Do the vanity first, then the tub or shower, then the toilet, then tile. Save each option as its own .blend file or collection ("vanity-60in-double," "vanity-48in-single") so a bad edit can't eat a good version. Ask it to label clearances.
6. Lock the cameras. Set two or three fixed camera views and render every option from the same spots. Side-by-side comparisons only work if the camera doesn't move.
7. Take the winner somewhere real. Bring your chosen layout and measured dimensions to Coohom or RoomSketcher for real catalog products, or straight to your contractor. They'll confirm what's behind the walls and what actually fits.

Used that way, an AI-built Blender model is a quick, cheap way to argue about layouts before anyone opens a wall. Just don't let it be the last thing that measures the room.