Generative interiors have a credibility problem. Type a prompt into most tools and you get a beautiful image of a room that could not exist: doors opening into sofas, windows halfway behind a bookcase, a rug the wrong shape for the floor beneath it.
The pictures are lovely. They are also useless for buying anything.
Understanding before styling
Mantha AI runs the problem in two passes. The first pass has nothing to do with aesthetics, it is a structural read of the space:
- Where the walls, floor and ceiling planes sit
- Which openings are doors and which are windows
- Where fixed services are: radiators, sockets, switches, vents
- What the circulation paths are, the space people actually walk through
Only once that map exists does anything get restyled. The geometry is a constraint, not a suggestion.
The rule we kept coming back to: a design you cannot walk through is not a design, it's a wallpaper.
Why circulation is the quiet hero
Most bad AI interiors fail on movement, not taste. A room can be gorgeous and still be unusable because there is no clean path from the door to the window. Human designers hold this instinctively; a model has to be told.
We reserve corridors around every opening and between major pieces, then let the generator work only in the space that's left. It rules out a lot of prettier compositions. It also means what comes out is something you could actually live in.
Materials that behave like materials
The second pass handles surfaces, paint, flooring, upholstery, tile. The trick here is lighting consistency. If the room has a north-facing window, a warm oak floor should pick up the cool cast of that light. Get this wrong and the render reads as a collage, even when every individual element is right.
We sample the room's existing light before applying anything, then relight the new materials to match. It is the least glamorous part of the pipeline and the one people notice most when it's missing.
What it is not good at yet
Being straight about this: Mantha is weak on rooms it can only partially see, and on heavily cluttered spaces where the structure is hidden behind belongings. Both are solvable, both are on the list, neither is solved today.
The point of all of it
A redesign is only worth anything if it survives contact with the actual room. Everything above exists to make one sentence true: what you see is what will fit.
The bit that took longest to get right
If there's one part of this pipeline that took more iteration than everything else combined, it's window and door recognition specifically, not because it's conceptually hard, but because the failure mode is invisible until it isn't. A misjudged wall angle produces a redesign that looks slightly off in a way people notice immediately. A misjudged door produces a redesign that looks perfectly fine right up until you realise the new wardrobe is now blocking it.
The fix wasn't a cleverer model, it was treating openings as harder constraints than everything else in the room. A door gets a wider "keep clear" margin than a plain wall does, because the cost of getting a door wrong (a room you can't actually walk into) is higher than the cost of getting a wall slightly wrong (a shelf that's a few centimetres off where it should be).
Why this matters more than it sounds like it should
None of this is visible in a finished redesign when it's working. You don't see the circulation map or the opening margins, you just see a room that looks like your room, restyled. That's deliberate. The entire structural-reading pass exists so that the part you actually look at, the styled result, doesn't need a caveat attached to it. "Ignore the sofa blocking the door" shouldn't be something you ever have to say about a redesign, and the amount of engineering behind that one sentence not being necessary is disproportionate to how boring the sentence sounds.
