Automating property photo editing

Some problems cannot be solved on the day of the shoot. The sky is grey, the lawn is scorched, the room is full of someone else's furniture, and the window is four stops brighter than the wall beside it.

Here is which model addresses which of those, and what each one will and will not do.

Match the model to the problem

The light was wrong. Photo enhancement corrects exposure, recovers blown-out windows and removes the colour cast from indoor lighting, while preserving the room's composition and furnishings. Run this first when you are chaining models, so everything downstream works from a clean input.

The sky was flat. Sky replacement is not a composite. It relights the scene to match the new sky, adjusting shadows and reflections across the whole photograph rather than pasting something new above the roofline. Six skies, from clear blue through to twilight.

The lawn was dead. Lawn enhancement fills bare earth, dead patches and worn tracks, and corrects drought-stressed grass to an even green. What it will not do is invent a lawn: the grass keeps its existing shape, no grass appears where there was none, and paving, gravel and borders stay as they are. Three intensities, from subtle to lush.

Something is in the shot. Object removal takes a plain-language list of what to remove — cars, bins, signage, cables — and rebuilds the surface behind it. There is no mask to paint and nothing to trace.

The room is furnished and the client wants it empty. Furniture removal clears the occupant's belongings down to a clean space, keeping the property's own fittings.

The room is empty and the client wants it furnished. Virtual staging, in any of sixteen styles.

The upsell that does not need subcontracting

Staging requests used to mean either turning the work away or sending it out. Running the model yourself means the turnaround is yours and so is the margin.

Worth being straight about what you are selling: this is a plausible version of the space, not a photograph of it. It is a visualisation, and in many jurisdictions it has to be disclosed as an edited image. Check the rules that apply to your listings.

Test it on your worst shoot

Output is watermarked by default and free to try — you pay only Replicate for the generation, on your own API token. So find the shoot you were least happy with and run it. That tells you more than any sample gallery, because the samples are always the easy cases.

When the output is good enough to deliver, a watermark-removal token is $100 a month and covers 1,000 images across all seven models. It renews monthly and an unused balance does not carry over, so it suits a steady workload better than an occasional one.

Batching a whole shoot

For one image, the Replicate playground is fine. For a package of thirty, call the models over the API — they run concurrently, so a shoot is a loop rather than an evening.