Room Design Visualizer — Repaint, Retile and Restyle Your Own Photo

Not for the result, and that is deliberate. Generative models drift in colour, so the same paint reference comes back a slightly different shade on each run — unusable if you are choosing an actual product. AI is used where it belongs: a segmentation model traces what you clicked. The picture itself is composited, so the colour is exact, the result is repeatable, and it renders immediately instead of queueing as a job.

A scene-parsing model — SegFormer trained on ADE20K, about 14 MB — labels every pixel of the photograph with one of 150 indoor classes. Clicking asks what the label is at that point and selects the whole region, so the answer is exact rather than a colour range that hopes for the best. It also means the tool can list what is in your room and let you pick from the list. Where two of the same thing touch, the Refine tool runs Segment Anything on a single click to separate them.

The ADE20K classes that occur indoors: wall, floor, ceiling, sofa, armchair, chair, stool, table, desk, bed, cabinet, wardrobe, bookcase, shelf, rug, cushion, pillow, curtain, blind, lamp, chandelier, sconce, painting, mirror, clock, vase, plant, television, refrigerator, radiator, fireplace, window and door. Anything it does not know can still be selected by hand with the wand or the brush.

Yes. Select it and press Remove, and the background is synthesised from the surrounding surfaces using multi-scale PatchMatch — the same family of algorithm as a content-aware fill. It continues walls, skirting boards and floors reliably. It does not invent objects, so it works best on furniture standing against ordinary surfaces rather than on something covering a bookcase.

Add a photo of the replacement. If it is a PNG with a transparent background, that edge is used as-is; anything else is cut out on your device. The old item is removed and the background rebuilt, then the new one is scaled to the footprint of what was there, corrected for the room's brightness and colour temperature, darkened slightly where it meets the floor and given a contact shadow cast away from the light in the photograph. Drag it to reposition it.

Because a real camera would not record it that way either. A surface photographed under warm indoor light picks up that cast, and rendering the raw swatch colour into a warm room reads as a pasted-on layer. The "match room light" control sets how much of the room's own illuminant is applied — turn it to zero to see the pure catalogue colour.

Raise "protect objects". It compares each selected pixel with the surface's own hue and releases anything too far from it, which catches sockets, switches, frames and radiators. Because it compares hue rather than brightness, the shaded half of a wall stays selected. Anything it misses can be removed with the eraser.

Yes. Drag the four perspective corners onto the real edges of the surface, then set its actual width in metres. Each material carries the physical size of one repeat — 300 mm for a metro tile, 1200 mm for an oak plank — so the pattern is laid at true scale and in true perspective rather than being stretched to fit.

No. The image is decoded into a canvas and every operation after that is arithmetic on pixels inside the page. The models are downloaded to your device rather than your photograph being sent to theirs, which is also why everything after the first load works offline.