What these models will not do

Most pages about AI image editing are a list of what is possible. This one is the other half, because knowing where a tool stops is what makes it usable.

They produce a plausible space, not a measured one

These are generative models. They produce a convincing version of a room, not a survey of it. Nothing they output should be relied on as a factual record of a property's condition, dimensions or fittings.

That matters most for renovation. Virtual renovation shows what a dated kitchen could look like modernised. It is not a quote, a specification, or a promise that the plumbing allows it.

Each model deliberately holds something still

The narrowness is the point, and each model draws its line differently.

Lawn enhancement will not create a lawn where there was none, will not change its shape, and leaves paving, gravel and borders alone. If you want the garden redesigned, this is not that.

Furniture removal keeps anything that would still be there for the next occupant — fitted kitchens, bathrooms, built-in storage. It clears belongings, not the property.

Virtual renovation keeps the room's layout and openings. It changes finishes and fixtures, at one of three scopes; it does not move walls.

Photo enhancement preserves the composition and the furnishings. It fixes how the photograph was captured, not what was in the room.

If what you need is outside those lines, chaining two models is often the answer — clear a room, then stage it — and sometimes the answer is that no published model does it, and it is a custom build.

Disclosure is your responsibility

Many jurisdictions require edited property photographs to be disclosed as edited, and some are specific about virtual staging in particular. We cannot tell you what applies to your listings, and the rules differ by market. It is worth knowing before you publish at scale rather than after.

A bad run happens

Occasionally a model produces something you would not publish. At one image at a time that is obvious and you re-run it. Built into an automated pipeline, it is not, so decide early whether a person approves output before it is published, or whether the people using your product can re-run it themselves.

Where the images go

When you run a model, you are using Replicate. The image you upload is handled under Replicate's terms and privacy policy, not ours. If image handling is a constraint for your business, read their policy, and talk to us if the models need to run somewhere else.

Try it and find the edges yourself

Output is watermarked by default and free to try — you pay only Replicate for the generation, on your own API token. Run the models on your own difficult images, not the easy ones. That is where you learn whether they fit.