Why does AI landscape design change my actual house?

Every AI yard app is one of two things, and the product page rarely says which.

The first is text-to-image. A diffusion model starts from pure random noise and denoises it step by step toward whatever a text description asks for. Your photograph, if the app accepts one at all, gets read once and converted into words — "a single-storey house with a lawn and a driveway" — and then discarded. Nothing from your original pixels enters the generation.

The second is image-to-image, often with inpainting layered on top. Your photograph is encoded into the model's latent space and becomes the starting point instead of noise. Denoising begins partway along the schedule, so the structure already present in the image survives to the end. AI Landscape Design, published by Skyble Lab, works this way.

Same category on the App Store, same screenshots, same promise. Completely different mathematics, and completely different results at the roofline.

Ask why a text-to-image yard app changes houses and the honest answer is that nothing in the pipeline is holding those pixels. There is no reference image, no conditioning signal derived from your photo, no loss term that penalises getting your dormer wrong. There is a text prompt and a random seed.

The model has never seen your house. It has seen millions of houses, and it samples a convincing average of them. That is not a defect — a latent diffusion model conditioned only on text is doing precisely what it was built to do, which is produce a plausible image matching a description.

"Plausible" is the operative word. A plausible house has a roof, windows and a door in sensible places. It has no obligation to have your porch column, your garage on the correct side, or your slightly off-centre entry.

So the output is not your property with new planting. It is a photograph of a house that does not exist, standing in a garden you also do not have.

In an image-to-image pipeline, one number decides almost everything: how much noise is added to your photograph before denoising begins. The Diffusers library calls it strength, and it runs from 0 to 1.

Set it low and generation starts close to your original. The house is safe and the yard barely changes — you get your own photo back with slightly greener grass and a few extra shrubs. Set it high and generation starts close to noise, which is where genuinely new planting comes from, and also where your roofline stops being a constraint.

The SDEdit paper that introduced this technique names the trade-off outright: realism and faithfulness to the input pull against each other, and the noise level is the knob between them.

There is no single value that gives you both across the whole frame. Which is why the tools worth using do not apply one value across the whole frame.

Inpainting is the way out of the strength trade-off. Rather than one noise level applied to the entire image, a mask marks which regions may be regenerated and which get copied back from the original at every denoising step, as the Diffusers inpainting guide describes. Beds, lawn, paving and gravel go inside the mask. The house, the fence and the sky stay outside it.

Mask quality then becomes the whole ballgame, and it fails in three directions.

A mask that stops short of the wall leaves a band of original siding the model has to blend into, which produces a smeared foundation line. A mask that overruns onto the wall gives the model licence to redraw brick, and it will. A mask feathered softly everywhere produces the failure homeowners describe as the house melting into the ground.

Getting that boundary right is a segmentation problem rather than a generation problem, and it is the part of these products that actually separates them.

A shrub can be redrawn wrong and nobody notices. A house cannot, for 3 reasons.

Straight lines. Diffusion models work in a compressed latent space, and long straight edges — a roofline, a gutter, a garage header — are what compression handles worst. A one-degree bend in a roof edge is glaring to a person and nearly invisible to the training objective.

Repetition. Windows sit at regular intervals, and a fourth window placed a foot closer to the third than the second reads as wrong even to someone who cannot say why. Human vision is unusually sensitive to broken periodicity.

Familiarity. People are extremely good at recognising buildings, in the same way they are good at faces, because both are things we have looked at all our lives. A homeowner has looked at this specific building every day for years. They spot a moved front door in under a second, and they stop trusting anything else in the image immediately afterwards.

Put your input photograph and the output render side by side on one screen and check 3 things, in this order.

The roofline. Trace the top edge of the roof from one end to the other. It should be the same line, at the same angle, meeting the neighbouring trees at the same points.

The window trim. Count the windows first, then check the gaps between them and the width of the frames. All of it should be identical, because none of it is part of a garden.

The front door position. Measure it against something fixed — the corner of the house, the edge of the driveway, a downpipe. A door that has drifted even slightly means those pixels were regenerated rather than preserved.

Any one of the three failing means the render is not of your property and cannot be used to decide what to build. Run this test on AI Landscape Design before trusting it, and run it on every competitor too.

Three fixes, in the order worth trying.

Re-shoot first. Stand back far enough that the whole house wall and the whole yard fit in one frame, at roughly chest height, in flat overcast light. A tight crop of grass gives the model no architecture to hold onto, and hard midday shadow gives it edges that are not real ones. Shoot in landscape orientation.

Then turn the ambition down. If the app exposes a strength, creativity or transformation slider, halve it. If it only exposes styles, choose one closer to what is already there — asking for a tropical courtyard when you photographed a snowy front lawn forces a large change, and large changes are where houses move.

If both fail, the app is the wrong kind and no setting will rescue it. A text-to-image model cannot be talked into preserving a building it never encoded. Move to a photo-to-photo tool instead.

Search "landscape AI" on either store and several different products come back from unrelated companies. "Landscape AI", "Landscapio AI", "Garden AI: Landscape Design", "Landscaping AI: Backyard Plan" and "Landscape AI – Garden Designer" are five separate apps from five separate developers. None of them are published by us, and several are general image models wearing a yard-shaped prompt.

AI Landscape Design is published by Skyble Lab. That is the only product in this article we have a stake in, and the only one whose behaviour we can describe from the inside.

This matters for the complaint at hand, because the apps that move houses and the apps that do not sit next to each other in the same search results under near-identical names. A one-star review saying "it changed my house" may well be about a different product than the one on your screen. Check the developer name on the store listing before writing off the whole category.