Skip to content
Room Stylist AIJoin the waitlist

How Room Stylist AI keeps your layout, and how we check it

Room Stylist AI asks an image model to restyle your room without moving the walls, windows or doors and then measures whether it did. A red trace is drawn from your photo and laid over the render. The render is rejected if less than 90% of the traced lines still sit on an edge. In our launch tests 17 of 24 room and style pairs passed.

By Sunny Patel. Last updated

The four steps

  1. Photo. You add one photo of the room and pick a room type and a style.
  2. Trace. A line detector (OpenCV's line segment detector) finds straight edges in your photo. We keep the ones that are the room itself: wall and ceiling junctions, floor edges, window and door frames, and built-ins. Furniture, curtains, cabinet door splits and tile joints are left out because they are meant to change.
  3. Restyle. For the examples on this site the photo went to Google's Nano Banana 2.1 image model (gemini-nano-banana-2.1) with an instruction to keep the camera position, walls, ceiling, windows, doors, floor and built-ins exactly where they are and change only the finishes and furniture for that room.
  4. Check. The trace is laid over the render. Every few pixels along each traced line we look for an edge in the render running the same way within 6 pixels. The share of traced length that finds one is the layout kept score.

What the score means

A photo scored against itself gives 100%. A photo of a different room scored against the same trace gives 29% to 50% because any room has some vertical and horizontal edges in roughly similar places. Passing renders in our tests scored 90.7% to 99.3%. We set the pass mark at 90%.

The score measures alignment of the room's main lines. It does not prove every detail survived: a socket, a radiator valve or a skirting board can still change without moving the score much. It also does not judge whether the render looks good. That part is up to you.

Before: small apartment kitchen with wood cabinets, a steel fridge, white floor tiles and a breakfast bar.AI generated modern farmhouse restyle of the same kitchen: white shaker wall cabinets, an oak base unit, white tile splashback, black tap and a jute rug. Fridge, cooker, dishwasher and worktop stay where they were.
AI generated exampleKitchen. Layout kept score 99.3%. Before photo: Kurt Kaiser, CC0, via Wikimedia Commons.Red lines: walls, windows and doors traced from the before photo. 99.3% of the traced line length still sits on an edge in the render.KeptWalls and ceilingWindow and door openingsFloor areaAppliance and worktop positionsChangedWall cabinets: white shakerBase unit: oakSplashback: white tileTap: blackRug: juteAccessories

Every test result

Before launch we rendered 8 real room photos in three styles each with up to three attempts per pair. 17 of 24 pairs passed and 12 of them passed on the first attempt. Each attempt costs the same. A pass on the third try costs three times as much to make.

Layout kept score per room photo and style, October 2026
Photo and styleBest kept scoreAttemptsResult
Kitchen 1, Scandinavian97.6%1Pass
Kitchen 1, Modern farmhouse99.3%1Pass
Kitchen 2, Japandi98.0%1Pass
Kitchen 2, Modern farmhouse90.7%2Pass
Living room 1, Japandi97.1%1Pass
Living room 1, Scandinavian97.5%1Pass
Living room 1, Boho98.0%1Pass
Living room 2, Japandi97.3%1Pass
Living room 2, Scandinavian95.0%1Pass
Living room 2, Modern farmhouse96.6%1Pass
Bedroom 1, Japandi94.3%2Pass
Bedroom 2, Japandi97.3%1Pass
Bedroom 2, Scandinavian95.0%2Pass
Bedroom 2, Modern farmhouse96.5%1Pass
Bathroom 2, Japandi99.1%2Pass
Bathroom 2, Scandinavian99.2%3Pass
Bathroom 2, Industrial99.1%1Pass
Kitchen 1, Japandi48.3%3Rejected
Kitchen 2, Scandinavian82.8%3Rejected
Bedroom 1, Scandinavian59.3%3Rejected
Bedroom 1, Boho59.9%3Rejected
Bathroom 1, Japandi79.4%3Rejected
Bathroom 1, Scandinavian67.8%3Rejected
Bathroom 1, Modern farmhouse72.1%3Rejected

Pass mark 90%. The attempts column shows the attempt that passed for passing pairs. All three attempts failed for rejected pairs and the best score is shown.

What goes wrong

Seven pairs failed all three attempts. The walls were not moved one by one in the cases we looked at. The model reframed the whole shot. A zoom of one to three percent or a small shift is enough to push every traced line off its edge. Our wide-angle bathroom photo failed in all three styles. The model straightened the bent lines and rebuilt the room from a new viewpoint.

Kitchen 1, Japandi: 48.3%

Before: small apartment kitchen with wood cabinets, a steel fridge, white floor tiles and a breakfast bar.Rejected AI generated Japandi render of the same kitchen: pale oak cabinets and a paper lamp, but the whole view has shifted up and to the side, so a doorway that was out of shot now appears on the right.
AI generated exampleRejected after 3 attempts. Before photo: Kurt Kaiser, CC0, via Wikimedia Commons.Red lines traced from the before photo. Only 48% of the traced line length sits on an edge in the render, below the 90% pass mark.

The model reframed the shot: about a 1% zoom plus a shift of roughly 50 pixels. The cabinets and fridge are still in the right order, but every wall and frame line has moved off its trace.

Bathroom 1, Japandi: 79.4%

Before: small tiled bathroom photographed with a wide lens, with a curved tiled bath, a wall basin, a mirror and a toilet.Rejected AI generated Japandi render of the same bathroom: plaster walls and white stacked tiles, but the camera has moved, the basin and mirror have shifted right and the bath has changed shape.
AI generated exampleRejected after 3 attempts. Before photo: Alf van Beem, CC0, via Wikimedia Commons.Red lines traced from the before photo. Only 79% of the traced line length sits on an edge in the render, below the 90% pass mark.

The wide-angle photo bends the tile lines. The model rebuilt the room with straight lines and a new viewpoint, so the basin, mirror and bath edge no longer sit where they were.

Your photo and privacy

Rendering is not switched on yet. Nothing is uploaded today because the upload box on each page only previews your photo in your own browser. Before uploads open, the privacy notice will say which model processes your photo, what it is used for and how long it is kept.

Where the example photos come from

The before photos are real rooms from Wikimedia Commons and are used under CC0, public domain or CC BY licences. Every after image is AI generated and labelled that way. Photo credits:

Who makes this

Room Stylist AI is built and run by Sunny Patel. Sunny is a UK SEO consultant and web developer trading as ND Media. Sunny is not an interior designer. The styles come from the image model. The checks here are about whether your layout survived.

Questions

Which AI interior design app keeps my walls and windows the same?

Room Stylist AI shows it. Every example carries a red trace of the original walls, windows and doors and a measured layout kept score. Renders below 90% are rejected.

What photo gives the layout check the best chance?

Shoot level from the doorway with the normal 1x lens. Wide-angle photos bend straight lines and our wide-angle bathroom failed in all three styles.

Why not just straighten a reframed render?

We tried aligning renders back onto the photo before scoring. It raises scores and it also hides the fact that the image you would see has moved. The score is taken on the render exactly as you would receive it.