How AI room redesign from a photo works

The short version

A room redesign is an edit of your photograph. It is not a new picture of an imaginary room, and it is not a 3D model. Kubikly analyses the photo, writes a design specification from your request, edits the photo to that specification, and checks the result against the original before showing it to you. If the room moved, it tries again.

Step 1: reading the room

Before anything is designed, the analysis identifies the room type and separates two kinds of thing:

• Fixed architecture: walls, windows, doors, ceiling, radiators and built-ins. These are the frame every result has to keep.
• Movable items: furniture, decor and clutter, each with where it sits in the frame. These are what can change.

That split makes everything after it possible. A tool that does not know which part of the picture is a window will happily move it when you ask for a new sofa. The list of movable items is also what you mark Keep, Remove or Replace, so you can say “keep the sofa, remove the boxes, replace the desk chair” and have it followed rather than guessed.

Step 2: a written design specification

Your request is turned into a structured specification before any pixels change. The request is made of a mode (Redesign, Clean & redesign, Declutter only or Full transformation), a style (one of 16 presets, your own description, or a reference photo of a room you like), a budget level, your per-item marks and anything you typed. The specification states the palette, the materials, the lighting, and what to keep, remove and add.

Two things follow. The result can say what it was asked to do: the spec panel shows what was kept, removed and added. And a revision changes the specification starting from the previous design, so when you ask for “cream bedding”, what you did not mention stays as it was.

Step 3: editing the photo

An image model then edits your photo to the specification. Editing is different from generating: the starting point is your room, with its light, its viewpoint and its proportions, so the result is still recognisably your room. Up to three options can be produced at once from the same analysis, for example three styles side by side, and because they share the analysis, the differences between them are the design, not a different reading of the room.

Step 4: checking the result

Image models drift. Ask for a new rug and a window can slide sideways or a door frame can quietly disappear. So every result is checked twice before you see it:

• An edge-geometry comparison against the original photo. Walls, windows and door frames make strong edges; if they no longer line up, the room has moved.
• An independent visual check: is it the same room, from the same viewpoint, with windows, doors and walls unchanged, the style applied and the result realistic?

A result that fails is retried, up to three attempts. If none passes, the job stops and the credits are refunded automatically, so a design that failed its checks costs you nothing.

What it is good for

• Deciding a direction before you spend: a paint colour, a style, whether the sofa is worth keeping.
• Seeing a cluttered room clear, which is often the push needed to clear it.
• Comparing styles on your own room instead of on a showroom's.
• Briefing someone else, whether a partner, a decorator or an interior designer, with a picture of your room rather than a mood board of other people's.

What it cannot do

• Measure anything. The result is one viewpoint with no dimensions, so it cannot tell you whether a sofa fits.
• Move walls, windows or doors. That is deliberate, and the checks enforce it.
• Show the room from another angle. The viewpoint is your photo's; another view needs another photo.
• Produce a shopping list. There are no product links, no furniture prices and no brand identification. The budget level steers the look, not a quote.
• Match every detail. Results are checked, but small things such as a socket or a picture frame can differ from the real room.

How it differs from a text-to-image generator

A text-to-image generator draws a new room from a prompt. It can be beautiful, and it is not your room: the window is wherever the model put it. Image-to-image tools start from your photo but usually repaint it freely, and each further request drifts further from the original. Room Redesign also uses an image model at its core, so the difference lies in everything around it: an analysis of what is fixed, a written specification, per-item marks, a geometry check against the original with retries, a refund when no acceptable result is possible, and revisions that start from the previous design.

How it differs from a 3D room planner

A 3D planner, including Kubikly's own floor-plan-to-3D studio, builds a model of the room with real dimensions. You can walk around it, measure it, move walls and export it, but it needs a floor plan or a lot of drawing. A photo redesign needs one photo and gives a photoreal picture of your actual room, usually in about a minute, but it is a picture: no measurements, one viewpoint, no walls moved. Use the photo redesign to decide how a room should look, and a 3D model of the home when you need to know whether things fit or want to change the architecture.

Getting the best result

The photo matters more than the prompt. A wide shot from a corner or the doorway, in daylight, with the floor and the ceiling line in frame, gives the analysis the most to work with. A close-up of one wall gives it little room context and is the weakest case. The photography guide covers this in detail, including what to take out of the frame for privacy.