A kitchen plan often has to show a door style that has not been ordered yet. The installation photograph is already approved: the L-shaped run is in the right place, the window sits over the sink, the appliance stack is the one on the drawing, and the worktop already has the stone the client signed. Then the client asks to see the same kitchen in oak shaker, in a flat white slab, in a dark green painted front, and in a fluted walnut — and no one is paying to hang four full sets of doors before Thursday's meeting.
If you generate a new kitchen from a text prompt, you usually get a handsome kitchen that is not this kitchen. An island appears. The window moves. The worktop changes because "kitchen" in the training data often has a different stone. That is the wrong job. The photograph already holds the layout. What you need is a front restyle, not a new room.
This is a practical workflow for doing that in the browser, before you rebuild the kitchen in a visualization scene or schedule a second site photograph you do not need.
Treat the layout as the contract
Cabinetry image-to-image only works if you are explicit about what must not move. The run, the openings, the worktop, and the appliances are the architecture. The door style, color, and handle language are the option set.
Write the split down before you generate:
- Keep: the L-shape or galley layout, the window position and size, the sink and tap, the worktop material and thickness, the appliance stack and their door lines, the plinth, the ceiling height, the camera height and angle, the crop, and any identifiably real irregularities (a slightly uneven end panel, a visible gable, a tile joint that already exists).
- Change: the cabinet front style and color, the handle or push-to-open language, and — if the brief really requires it — the finish of a specific secondary front such as a tall larder. Not a new island "to make the room more current." Not a different worktop. Not a moved window. Not a new tile because the wall looks calmer without the real one.
If the brief quietly includes "make it more high-end" or "give it a brochure look," the model will often "improve" the kitchen. That improvement is how a real installation becomes a cousin of itself: a wider window, a second sink, a hidden appliance stack, a worktop that was never specified, a ceiling that has grown a rooflight.
One change family per run is enough. Do not ask for new oak shaker fronts, a new worktop, and a new floor in the same prompt. Fronts first, on the original worktop and the original light. Stone or flooring later, from a keeper that still matches the layout.
Start with a reference that can survive a restyle
Not every finished-kitchen photograph is a good source. A shot from too close in a corner, with heavy wide-angle stretch, or a detail of two doors with the rest of the run out of frame, will not give the model a stable layout to re-front. You want a still in which the whole relevant run is readable and the openings are countable.
A usable source usually has:
- A clear view of the main run and the tall units, so you can count door modules. If you cannot number the fronts on the source, you will not notice when the output adds one.
- A readable worktop edge and a visible window. Those two lines are how you catch a layout drift.
- A quiet enough foreground. A person, a dog, or a fully styled breakfast spread will be redesigned when you only asked for new doors.
- A single lighting condition. Keep the existing window light unless you are doing a separate lighting job. Mixing "new fronts" with "golden evening" is how the room becomes a different room.
This is not a whole-room restyle, and it is not a furniture-in-context exercise. You are not asking to move the kitchen into four other spaces, and you are not asking to relight the architecture. Those are different jobs. Here the room stays, and only the fronts change.
Write the prompt as a change list, not a mood board
"Make a luxury modern kitchen, cinematic interior, photoreal, 8k" is a mood board. It invites a new kitchen. A change list names what stays and what moves.
A prompt that has worked for this job looks more like:
Same kitchen and same layout as the reference. Keep the exact cabinet run, window, sink, worktop, appliances, ceiling height, crop, camera angle, and lighting. Change only the cabinet fronts to oak shaker doors with a clear finish and simple bar handles, on the same modules. Do not add an island. Do not change the worktop. Do not move the window. Do not redesign the room.
Name the front in joinery language, not in brochure language. "Warm contemporary oak kitchen" is how you get a generic villa. "Oak shaker doors, clear finish, same modules, simple bar handles" is closer to a door sample.
Name the surfaces that should not change, and stop there. If the worktop, the plinth, and the end panels should stay as installed, say so. Models like to "complete" a material story by wrapping the new oak around the worktop and the floor.
Avoid words that sound like a redesign: open plan, add island, bigger window, hidden appliances, luxury renovation, cleaner lines. Those are invitations to change the layout.
A browser pass is enough for a first door-style set
You do not need a local kitchen visualizer to try four front styles on a layout you have already photographed. A browser-based Bild-zu-Bild-Generator is enough for this kind of restyle: upload the approved installation still, write a change list that protects the layout and the worktop, generate a short set of variants, and download the ones that still look like the same kitchen. Multiple models are available in the same workspace; if one pass invents an island, try a different model on the original still rather than stacking a "fix" on a room that has already drifted.
Stay conservative with file prep. A sharp photograph of the approved kitchen is better than a wide-angle phone snap that already stretches the end cabinets, and better than a heavily styled magazine crop that hides the tall units. If the source is within a normal upload size, do not pre-upscale it in a way that invents wood grain the camera never saw. Generate a small batch — a handful of variants is enough to see whether the window and the appliance stack are holding — and pick from those before you spend another pass.
The point of the browser pass is not to replace a manufacturer's door sample or a planning-quality visualization. It is to hang four comparable options in a meeting while the order is still open.
Check the window, then the modules, then the worktop
Review in a fixed order, or you will approve a pretty oak kitchen that is slightly the wrong kitchen.
- Layout first. Is the window still over the sink, in the same size? Is the run still an L, or a galley, as installed? If an island has appeared, or a wall has opened, discard the frame.
- Modules and appliances. Count the doors and the drawers. The oven, the dishwasher, and the fridge should still sit where they sat. A "cleaner" stack that hides a real appliance is a failed restyle for a client who already owns those machines.
- Worktop and splash. The stone, the thickness, and the upstand or tiles should not have changed unless you asked. A new marble is a different quote.
- Front language. Shaker should still have a rail and a stile; a slab should still be a slab; fluted fronts should have a plausible pitch, not a printed texture that falls apart at the handle. Handles should be consistent and should not multiply.
- Light last. Do the window and the worktop still receive light from the same side as the original still? A second sun on a newly heroic island — especially an island that should not exist — is a common giveaway in a four-up comparison.
If only a handle finish is wrong, you can usually run a second, narrower pass from the best variant. If the layout has already drifted, go back to the original still. Do not stack restyles on a kitchen that is already a little bit wrong. Each generation will "help" it further.
A simple four-front set from one still
A practical set for a client meeting, all from the same approved installation photograph:
- Existing fronts, proof pass. Run the source through the same prompt structure with "keep the existing cabinet fronts" as a control. If the model cannot reproduce the kitchen it was given, do not trust it with fluted walnut yet.
- Oak shaker. Clear or lightly stained oak, a true shaker profile, simple handles, same modules. Watch that the model does not also replace the worktop with a butcher block "to match."
- Flat white slab. A painted or lacquered slab, with either the specified handle or a push-to-open front if that is the brief. Watch for a sudden high-gloss that the client did not ask for, and for missing grain only where you expected a paint.
- Dark green paint and fluted walnut, as two later options. Dark green is a good third pass because it will show immediately if the modules have moved. Fluted walnut should keep a buildable flute pitch and should not turn the tall units into a different, more monumental mass.
Keep the crop and the camera height constant across the set. A client comparing four images side by side will notice a zoom-in or a slightly wider lens faster than they will notice a slightly different oak grade.
When a local model is the better tool
Graduate to a visualization scene when the fronts must match a manufacturer's exact door program, when you need the same kitchen under a locked sun path, or when the output will be used as a construction drawing rather than as a meeting option. Until then, a careful source still, a written keep/change list, and a short browser series will usually get you a usable front set without another site visit.
The test is simple: could the kitchen planner still recognize their layout? If yes, the restyle is doing its job. If you have to explain that it is "the same kitchen, we just opened the room a little," the plan has already moved, and you should start again from the original still.

