AI photo editing for property listings, and where the line is
What these tools change, what they must not change, what Australian law requires, and what this product does and does not do about it.
What AI photo enhancement changes, and what it must not
AI real estate photo enhancement is the use of a generative model to edit a photograph of a property before it is advertised. At its narrowest it does what a professional retoucher has always done: correct exposure, white balance, colour and sharpness, so the photograph looks like the room looked. At its broadest it can add furniture to an empty room, replace an overcast sky, or place people in a scene.
The line that matters is not how much the image changed. It is whether what changed is part of the property. Furniture, clutter, cars, weather and people are movable or temporary, and none of them is what a buyer is purchasing. Structure, surfaces, materials, fixtures and fittings are. An edit that brightens a kitchen is a correction. An edit that removes a crack from a wall is a misrepresentation, however subtle it looks.
RealEstateKit.ai is built on that distinction. The tools may change what is in a room and what is in the sky. They never change the property.
What each tool does
Photo editing runs as a chain of separate, named tools rather than one undifferentiated filter. Each edit is one tool applied to one image, recorded in order, and reversible. The agent chooses which tool runs and with what settings.
Enhance corrects exposure, lighting, colour balance and sharpness. On interior shots the strength of that correction is the agent's choice, with lighting and colour tone selectable per image; where no tone is chosen the instruction sent is to keep colour temperature neutral and balanced. Exterior shots take a different instruction set entirely, asking for natural enhancement of colours and sky rather than a colour grade. Every enhance, on either path, also carries a closing instruction that beyond the enhancements explicitly listed, nothing may be added, removed, moved or replaced, and no new item may appear in the scene.
Beyond the correction, Enhance offers four optional settings, all off by default and applied only when the agent selects them for a particular image: replacing a dull or overcast sky, boosting the colour of existing grass and planting, cleaning surface soiling from a building exterior, and adding people to a scene for lifestyle imagery. The first three apply only to exterior shots.
Cleaning a facade is where the distinction above does real work, so the instruction is specific. The tool is told to remove surface soiling only, meaning dust, dirt film, grime, cobwebs, bird droppings and weathering deposits sitting on top of the surface. It is told not to remove, repair, cover, lighten, blend away or reduce cracks, chips or spalling, rust or rust run marks, mould, mildew, algae or lichen, water staining, damp or efflorescence, peeling, flaking or faded paint, or missing or damaged cladding, render, mortar, roof tiles or guttering. Those are the condition of the property, not dirt. And where the model cannot tell which it is looking at, it is told to leave the mark unchanged.
Declutter removes movable contents, from benchtop clutter and cables through to furniture, and depending on what the agent asks for it can leave a room bare. It is instructed to subtract only: the prompt forbids adding any element not present in the source, and asks for a result containing fewer items than the source, never more. Structure, built-in cabinetry and fixtures are to be kept identical.
Stage furnishes an empty or sparsely furnished room, the digital equivalent of physical property styling. It adds furniture and decor, and it can also replace furniture already in the room, so a staged photograph may not contain the pieces the room started with. It is instructed to add furniture and decor only, and not to invent fixtures, structural features, appliances, plants, signage or scenery.
Across all three, two constraints apply regardless of tool. The property's structure, surfaces and materials are to be preserved exactly, with the only permitted changes being those in the tool instruction. And no text may be added, invented or altered: no letters, numbers, logos, brand names, watermarks or signage, and where text in the original is blurred or illegible it is to stay that way rather than being made legible.
Where these tools go further, named plainly
Three of the things this product does go beyond correcting a photograph. Each is defensible. None is defensible silently.
Video clips insert people, and that is what happens in practice
The video feature turns a still photograph into a short marketing clip. It does this in two stages. First a generative model produces a new still from the agent's photograph, adding people to the scene, and it is that generated still, not the photograph, which then becomes the video. The scenes on offer are a family, a couple, one person, a group of friends, and one option that adds furniture rather than people.
Every option but the last inserts people, and the one that inserts a family is the default. The clips ship in the export pack.
The generated still is shown to the agent before the video is rendered, and the render only starts once the agent approves it. That approval is the only approval gate in the product, and it is worth being precise about what it covers: it gates the generated image before the expensive render, after the photograph has already been sent to the model. It is not a check on the photograph, and it has no equivalent anywhere in the photo tools.
The generated still itself never ships. It is kept against the clip record and does not enter the image set, so what reaches the export pack is the rendered clip and nothing else.
The instruction sent with that generation says that adding people, soft furnishings and decorative items is the legitimate purpose of the edit, and that adding anything else is forbidden: no structural element, appliance, fixture, plant, signage or scenery beyond those. It carries the same structural preservation and the same prohibition on inventing text as the photo tools.
Where the export pack contains clips, it carries a disclosure stating that the clips are AI-generated from still photographs, that the motion and any people shown were digitally created and were not present when the photographs were taken, and that the people shown are not residents, occupants or owners.
Virtual staging shows furniture that is not there
Staging adds furniture and decor to a room, and can replace furniture already in it. It is the digital equivalent of physical property styling, which agents have done for decades, and the argument for it is the same: an empty room is hard to read, and furniture lets a buyer picture themselves in the space. What makes it legitimate is that none of it is part of the property. What makes it misleading is publishing it without saying so.
Where a staged image is in the pack, the pack's disclosure names it, states that staged images show digitally added furniture and decor, and says those furnishings are not included with the property.
Sky replacement changes the weather, not the property
Replacing a dull or overcast sky is the most common cosmetic edit in property marketing and the least consequential, because weather is not a feature of a house. It is still a change to the photograph, and an agent who has never looked at the original should know it happened. It applies to exterior shots only, is off by default, and runs only when the agent selects it for a particular image.
And one that exists but has not been used
Enhance can add people to a photograph, in the same way the video pipeline does. It is off by default, and unlike the other optional settings it is not restricted to exterior shots. It has never been used in production: no image in this product has ever been edited with it.
It is described here rather than quietly omitted, because a page about what a product can do should not be a page about what it happens to have done.
What Australian law requires
Two provisions of the Australian Consumer Law do most of the work here, and they apply nationally. Section 18 prohibits misleading or deceptive conduct in trade or commerce. Section 30 applies those rules to the sale of land, and prohibits false or misleading representations about the characteristics of the land, its location, the price payable, and the facilities associated with it. Breaching section 30 carries a maximum penalty of $220,000 for an individual and $1.1 million for a company. In New South Wales, section 53 of the Property and Stock Agents Act 2002 also allows a person to claim damages for misrepresentation or concealment in a sale or purchase, whether or not an agency agreement has been entered into. Other states have their own agent legislation; the Consumer Law provisions apply everywhere.
NSW Fair Trading has published guidance on what this means for photographs specifically, and it is the most direct regulator statement on the subject in Australia. It says four things that matter to anyone using these tools.
- Agents must not allow photographs to be modified so that they no longer truthfully and fairly represent the property, and must not change how a property appears by digitally removing or adding features. Adjusting lighting to compensate for poor lighting conditions may be acceptable.
- A representation is misleading if it produces a reasonable belief in a state of affairs that does not exist. It does not matter whether anyone intended to mislead. What matters is the impression a reasonable buyer takes away.
- Silence can mislead. The guidance says plainly that representations can be misleading by omission. A staged photograph that is accurate in every pixel can still mislead if nothing tells the buyer the furniture is not real.
- Labelling is the remedy the regulator itself proposes. Its worked example concerns location shots: a photograph of a beach or parkland not taken on the property should be labelled as such, while a photograph taken on the property needs no label. The principle generalises. Where an image could leave the wrong impression, say what it is.
One more point from the guidance is worth repeating because agents sometimes assume otherwise. An agent cannot avoid liability by arguing that the buyer should have made their own enquiries. The agent is responsible for the representations they make.
Where these tools sit against that
Enhance's core correction is the closest thing here to the lighting adjustment the guidance treats as potentially acceptable, and it is the least exposed.
Declutter and Stage change the contents of a room, not the property. That is a real distinction and we think it is the right one, since a buyer is not purchasing the sofa. But an unlabelled staged photograph is exactly the case the guidance's omission line reaches, and no argument about what furniture is will help an agent who published one without saying so.
Sky replacement changes the weather. Weather is not a characteristic of the land.
Cleaning a facade is the closest call, and it is why that tool is narrowed the way it is. Removing dirt from a wall is not the same act as removing a crack from one. The first is cleaning; the second is concealing a characteristic of the property, and it is squarely what the guidance prohibits. The instruction sent to the model draws that line explicitly, which is not the same as guaranteeing the model honours it.
Video clips insert people and motion that were never there. Nothing in the guidance addresses AI-generated video, because it predates it. The principle does not need updating to cover it: the question is what impression the clip leaves, and a clip that appears to show a real family in a real house on a real afternoon leaves an impression that only a label corrects.
Sources: NSW Fair Trading, advertising guidelines for property agents, and the Australian Consumer Law, Schedule 2 to the Competition and Consumer Act 2010.
What to do about it
None of this is complicated. Four things cover it.
- Disclose virtually staged images in words. The caption field, the image label, or the listing description, wherever the portal shows text beside the photograph. "Virtually staged" is enough. Do it on every staged image rather than the first one, so there is no question about which is which. This product does not put a watermark or any marking on the image itself, so if the words are not in your listing they are nowhere.
- Disclose AI-generated video in the caption or description. A clip that appears to show people in the property is the strongest impression anything here creates, and it is the one most likely to be taken literally. Say that the clip is AI-generated and that the people in it are not residents.
- Keep the originals and compare them. Every edit in this product retains the untouched original, and both ship in the export pack. Looking at the pair takes seconds and is the only check that exists on what a model actually returned. It matters most on exteriors, where the difference between a facade that was cleaned and one that was quietly repaired is the difference between a correction and a misrepresentation.
- If in doubt, do not publish it. An edit you would have to explain to a buyer at an inspection is an edit that should not have gone out.
The export pack is built to make the first three easy rather than to do them for you. It contains a disclosure file naming what was generated with AI assistance, including a section covering video clips where the pack contains any, and a manifest recording which tools were applied to which image in the order they ran. Those are material to disclose with. They are not disclosure, because nothing in this product reaches your listing. You publish it, so the representation is yours.
That last point is also what the terms of service say: if you use virtual staging, disclosing that images are digitally staged is your responsibility. This page is not softening that. It is explaining why the obligation sits where it does, and giving you what you need to meet it.
What is applied, and what is kept
Four things make the difference between a product that says the right thing and one that is built to make it checkable.
The tools are separate and named. Editing is not one undifferentiated filter with a strength slider. Enhance, Declutter and Stage are distinct tools with distinct instructions, and each edit is one tool applied to one image. An agent who applied Stage to a photograph knows they did, and so does anyone reading the manifest afterwards.
Each tool carries explicit written constraints. Those constraints are quoted in the tool section above rather than summarised, because a constraint you cannot read is a claim rather than a constraint. The property's structure, surfaces and materials are to be preserved exactly, with the only permitted changes being those in the tool instruction. No text, letters, numbers, logos or brand names may be added, invented or altered, and text that is illegible in the original stays illegible. Enhance may not add, remove, move or replace anything beyond what was listed. Declutter may only subtract. Stage may add furniture and decor and nothing else.
Free-text instructions are screened before they reach the model. An agent can add their own instruction to an edit, and that text is checked against a list of prohibited request patterns: removing or deleting permanent structure and fixtures, changing layout or dimensions or extending the property, making a space look larger than it is, faking a view, hiding neighbouring properties, adding features that are not there, and adding text, logos or branding. On the photo path the check runs twice, once in the application before any record exists and again in the editing pipeline before the model is called. On the video path it runs once, in the pipeline. A blocked request produces a failure rather than an edit, and is never sent.
That screen is a keyword check rather than a semantic one, and it is worth being exact about what that means. It matches phrasings, so it catches the common ways of asking and does not catch every possible way of asking. It is a guard against the obvious, not a filter that understands intent. What sits behind it is the constraints above, which apply to every edit whether or not anyone typed anything.
Framing is taken away from the model where it can be. Before the first edit in a chain, the source image is cropped to the nearest aspect ratio the model supports, using ordinary arithmetic rather than a model, and that ratio is then sent explicitly. The effect is that the model receives a canvas it exactly fills, which removes the freedom to extend the scene beyond what the photograph contains. The same crop runs on the video path. It applies where the image dimensions can be measured and the crop succeeds; where either fails, the uncropped image is used instead, and that is the one condition under which the model has room to invent.
Originals are kept, and both ship. The original is the untouched stored file, not a derivative and not a link to wherever the image came from. Every edited image travels with its original in the export pack, paired by index in the same folder, so a comparison needs no extra step. Images that were never edited ship once.
Every edit is recorded in order. The manifest in the pack records which tools were applied to which image, in the order they ran, along with the aspect ratio the source was cropped to and any edit that failed. Reverting an edit flags it rather than deleting it, and a reverted edit cannot reach an export.
What this product does not do
Everything above describes instructions given to a model and evidence kept afterwards. Neither is the same as verification, and the gap between them is worth stating rather than leaving for someone to find.
Nothing checks what the model returns. Every safeguard described on this page sits on the input side: the tool constraints, the screen on free text, the crop that fixes the canvas. Once an image comes back, nothing compares it to the original, nothing measures what changed, and nothing blocks it. There is no verifier and no automated review. That is why this page says the tools are instructed rather than that they are prevented, and why the original is kept: the comparison exists, it just has to be made by a person.
Instructions are not always followed. Two occasions are on record where a model returned fabricated brand text despite being told not to invent text, and one where it extended a scene beyond what the photograph contained, inventing a ceiling and a floor. All three happened when the model was given a canvas that did not match its source, which is the defect the deterministic crop was built to remove. A subsequent run of six generations with the crop in place produced no fabricated text. Six generations is evidence that the fix helps. It is not evidence that instructions are always followed, and no measured rate of compliance exists, because none has been taken.
There is no approval step for photographs. A photograph can be enhanced, decluttered or staged and go straight into the export pack without anyone looking at it. There is a batch action that applies the default enhancement to every image at once, and the review step can be skipped, so it is entirely possible to export a set of edited images that no person has examined. The only approval gate in the product is on the video path, and it covers the generated still after the photograph has already been sent to the model, not the photograph before.
Nothing in the product reminds you to disclose. This page tells you to label virtually staged images and AI-generated clips. The product does not. No prompt, no checklist, no warning at export. The pack contains a disclosure file and a manifest, and it is entirely up to you whether either informs what you publish.
Nothing marks the images themselves. No watermark, no embedded notice, no metadata flag a portal would read. An edited image is a JPEG like any other. If a disclosure is not in your listing text, it does not exist.
The free-text screen is narrow. It matches phrasings rather than meaning, so a request it is designed to stop can be reworded past it. The constraints on each tool are what actually apply to every edit.
One capability is unmeasured. The facade cleaning tool was narrowed to surface soiling only, with an explicit list of things it must not touch, including cracks, rust, mould, water staining and failing paint, and an instruction to leave a mark alone where it cannot tell dirt from damage. Whether the model honours that narrowing on a real damaged facade has not been tested. The instruction is specific and recent; the outcome is not known.
That list is deliberately unhedged. Every item on it could have been written to sound smaller, and every one of them would have been less useful.
Common questions
Last updated: 21 August 2026