How AI Redesigns Your Room in 90 Seconds (and what's actually happening under the hood)

If you've ever stared at a beige living room and thought there's a better version of this somewhere — I just can't see it, you already understand the problem AI interior design solves.

The gap between a photo of what is and a rendering of what could be used to cost you a designer, a week, and a few hundred dollars. With Architerior AI it's a drag-and-drop, 90 seconds, and your browser tab.

This post walks through what actually happens in that 90 seconds — not the marketing version, the real pipeline.

1. The upload (0–3 seconds)

You drop a photo. In the browser we run a couple of boring but important steps before anything hits the server:

• Format normalization. iPhone HEIC, obscure TIFF variants, and oversized PNGs get transcoded to JPEG. Most AI vision models only speak jpeg / png / webp / gif — sending them a HEIC is an instant failure.
• Downscale if huge. Serverless request bodies cap at 4.5 MB on most platforms and Safari's fetch sometimes rejects very large JSON payloads with a cryptic "string did not match expected pattern" error. We resize to 2048 px on the long edge before we leave your browser — the vision model resamples to ~1568 px internally anyway, so you lose nothing.
• Context you typed. A sentence like "master bedroom, keep the hardwood floors, prefer warm neutrals" gets bundled with the image. Free-text context is the single highest-impact thing you can give an AI — more than any slider or dropdown.

2. The read (3–15 seconds)

The image + your context go to Claude. We don't ask it to "redesign the room" — that would hallucinate. We ask it to describe the room in structured JSON: dimensions, light quality, materials present, architectural features, obstructions. It returns something like:

{
  "estimatedDimensions": "12 × 14 ft",
  "ceilingHeight": "9 ft",
  "windows": [{ "wall": "east", "width": "4 ft" }],
  "flooring": "mid-tone oak hardwood",
  "constraints": ["preserve hardwood", "warm palette"]
}

That JSON is the ground truth every downstream step is pinned to. It's why your hardwood floors don't turn into marble between "upload" and "render" — the downstream prompt explicitly forbids changing what the structured read observed.

3. The reimagining (15–75 seconds)

Now the fun part. Given that structured read plus your context, we ask the model to propose three distinct interior-design visions — not three variations of the same thing. Each comes back as a paragraph-length design brief:

• Vision 1: "Warm Scandinavian — bleached oak, cream bouclé, black accents"
• Vision 2: "Japandi — low-slung walnut, linen, paper lantern"
• Vision 3: "Soft industrial — charcoal walls, brass fixtures, vintage leather"

Each brief goes to a different image model (we cascade through Flux Kontext Pro → ControlNet → a Gradio fallback, picking whichever is healthy when your request lands). Kontext Pro is the preferred path because it's an image-editing model, not a text-to-image one — it takes your actual photo and rewrites it rather than generating something new that resembles it. That's why the windows stay where the windows are.

4. The delivery (75–90 seconds)

The three renders come back as raw image bytes, stream into your browser, and persist to your Supabase bucket tied to your user ID. You get a gallery. You pick a favorite. You share it, or you upload a different room, or you open the precise DXF pipeline and start over with dimensions.

What this isn't

A confession worth making: this isn't photogrammetry. The AI doesn't measure your room — it estimates. If exact dimensions matter (you're actually quoting a remodel), you should upload a floor plan or a DXF, not a phone photo. The DXF pipeline is millimeter-precise; the photo pipeline is design-precise.

It also isn't buildable. A rendering shows you what a space could feel like. A contractor needs a plan. Use the rendering to sell the vision; use the floor-plan analyzer to actually cost it.

What's next

We're working on a couple of obvious extensions: iterate-on-render (change the couch, keep everything else), before/after sliders for the gallery, and a "save to dashboard" that actually uses the database we built instead of browser session storage. If any of those sound like a must-have, let us know — we prioritize based on what readers ask for.

In the meantime, if you haven't tried it yet, upload a room photo and see what happens in 90 seconds. The first three redesigns are free, no credit card.