AI and real estate photos: how we showed a house that wasn't finished
In April 2024, for Engel & Völkers Genova, we photographed a house that was still a building site and finished the pictures in post-production. What Content Credentials record, how we label a mixed picture, and where AI does not belong.
By Leo de Bruyn, founder of InStudio
The lawn in the picture above was not there when we took the photograph. In its place was a rectangle of bare earth with site debris around it. We generated the grass in Adobe Firefly, and the original file records it.
The picture comes from a shoot we did in April 2024 for Engel & Völkers Genova, the Genoa office of Engel & Völkers. We ran the agency’s content from April 2024 until early 2025. The house was still a building site, and the agency needed buyers to see it finished. Here is how we did that with a camera and post-production, what Content Credentials are, how we label a mixed picture, and where we think AI does not belong in property marketing.
The house as we found it
The terraces were bare earth. In the yard below the house stood two diggers, a pile of gravel, building materials and parked cars. A photograph of that proves the house exists. It tells a buyer very little about living there.
So we photographed the house as it was, from the ground and with a drone, and finished the pictures in post-production. In the drone view, the yard gained a lawn, a long pool and sun loungers, and the building site disappeared. In the view from straight above, the terraces became lawns, with a pool in the yard and a second one beside the house.
Two pictures finished with Adobe Firefly
For the terrace beside the house, we used generative AI. In Adobe Firefly we made two versions of the same frame, one with a lawn and one with a pool. The house, the railing and the sea are as we photographed them. The bare earth became grass or water, and the clutter around it is gone. Through the open doors you can still see packing boxes inside: we finished the ground, not the house.
Both original files carry Adobe Content Credentials, which record the use of Firefly. On the other finished views of the house, the slider says “Finished in post-production” and names no tool. We name a generative tool in a caption when the file records it.
What Content Credentials are
Content Credentials are a record of where a file comes from and what was done to it. They follow an open standard published by the Coalition for Content Provenance and Authenticity (C2PA), whose steering committee includes Adobe, the BBC, Google, Microsoft and Sony.
In the standard’s own terms, the record is a set of statements about the file, signed by the software or device that did the work, and usually embedded in the file itself. If anyone alters the record, its hashes no longer match, so the change shows (C2PA explainer).
Adobe says that Firefly outputs include Content Credentials indicating that AI was used. The Content Authenticity Initiative, which Adobe founded in 2019 and still leads, points to its Verify website as a place to read them.
Two limits matter. First, the C2PA says plainly that Content Credentials make no judgement about whether a file’s provenance data is “true”. They show what was recorded about how a file was made, and whether that record is intact. They do not say whether the picture is fair. Second, the record can be removed from a file. The standard describes ways to recover it, through a watermark or a fingerprint of the content and a copy of the record stored online. Our own case shows the problem: the resized copies of these two pictures on our site no longer carry the record.
How we label a mixed picture
Because metadata can go missing, we write the label in words, next to the picture. On our site, every picture that mixes real photography and AI says so in text, in its caption or in the sentence beside it, so a screen reader reads it out too. Our before-and-after slider names both sides: “As photographed” on one, and “Finished with Adobe Firefly” or “Finished in post-production” on the other. And each client knows exactly where AI was used in their work.
If you publish pictures like these in a listing or on a social account, three habits help. Keep the original file with its Content Credentials, along with the photograph as it was shot. Write the label as text beside the picture, in the language of the listing. And name what was added rather than writing just “AI”: “lawn generated”, “pool added in post-production”.
Where AI helps, and where it does not belong
Most questions about AI-generated real estate photos come down to one: will the buyer feel misled on the visit?
Generative AI is good at showing what could be, such as a garden on ground that is still bare earth. It is the wrong tool when a picture stands as a record of what exists. For property, that means a buyer should never discover on a visit that a picture altered what is already there: the view, the neighbours, the size of a room, the state of a wall. Finishing a garden on a building site is one thing. Widening a terrace or removing the building next door is another.
Leave invented people out too: a terrace does not need a generated stranger on the sun lounger to sell.
When a camera gets the shot, we use the camera. Each month, for Engel & Völkers Genova, we photographed and filmed the villas and apartments the agency was selling: living rooms, kitchens, bedrooms with the sea in view. In the case study, only these two pictures are labelled as using generative AI.
Rights and consent
Ask three questions before any mixed picture. Whose photograph is it? Footage and photographs you give us stay yours. Who is in it? We do not generate a real person, their face or their voice, without that person’s consent. What was the generator trained on? Adobe says its Firefly models are trained on licensed content from Adobe Stock and on public domain content where copyright has expired, and that it does not train on users’ personal or generated content (Adobe). If you commission AI imagery, ask your studio which tool it uses and what the tool’s maker says about its training data.
Two more checks complete the list. Make sure the owner of the property, or whoever runs the building site, has agreed to the shoot and to the pictures being finished. And ask your legal team to check the advertising rules wherever the listing will run.
What it costs, and what moves the price
People often look up the cost of AI video production as if AI replaced the shoot. In our work it rarely does: our generative work usually starts with a camera. So the price has two parts, the shoot and the finishing.
What moves it: how many pictures, or seconds of film, need finishing; how many versions of each (we made two of the terrace); whether a drone is needed; and how many rounds of review your team needs. The examples in this article are stills. In a film, every frame of a finished shot has to agree with the next, which means more checking.
We will not put a figure here, because a fair one depends on the brief. If you have a property, a product or a place to show before it is ready, tell us about it, or read how we work in AI video production and content creation.