Virtual staging with ChatGPT, Gemini or Canva: what works and what breaks

How we looked at this

This guide is written from what OpenAI, Google, Canva and Adobe publish about their image tools, read on October 22, 2026, and from the staging and MLS pages our other guides already cite. We did not run the same room through ChatGPT, Gemini, Canva and Photoshop and score the results, so there are no quality rankings here. Where a company's page would not load for us (OpenAI's help center and usage policies, Canva's help pages all returned 403), we say so and use the pages that did.

What the general image models actually do

ChatGPT, Gemini and Canva's AI tools are general image models. You give them a photo and a sentence, and they produce a new image that matches both. Google's API documentation describes the job as using "text prompts to add, remove, or modify elements, change the style, or adjust the color grading"; its August 2025 post on the Nano Banana editing model says it can "alter specific parts of an image while preserving the rest." OpenAI's developer documentation describes an edits endpoint that can "edit existing images, generate new images from reference images, and edit parts of an image using a mask." That is the whole pitch: describe the furniture, get the furniture.

The part the pitch leaves out is that the model regenerates the picture rather than pasting furniture onto your pixels. Four consequences follow, and all four are documented by the companies themselves.

Usage terms point the same way. Google's Generative AI Prohibited Use Policy, last modified December 17, 2024, says "Do not engage in misinformation, misrepresentation, or misleading activities," and lists "Misrepresenting the provenance of generated content by claiming it was created solely by a human." A staged room presented as a photograph of the furnished home runs into both. OpenAI's usage policies would not load for us (403), so we cannot quote them; and Canva's AI product terms did the same.

ChatGPT, Gemini, Canva and Photoshop, one by one

ChatGPT (OpenAI's image models). Upload a photo and describe the staging in the chat. OpenAI's current developer documentation names the gpt-image-2.5 models and recommends the three output sizes above. Editing is prompt-based, with an optional mask that the model treats as a hint rather than a boundary. OpenAI's help-center article on creating images in ChatGPT returned 403 to us, so plan limits and the consumer-side rules are not quoted here.

Gemini (Nano Banana). In the Gemini app you can "upload an image from your device that you want to edit," and editing "isn't available to users under 18." Downloads are 2K with a Google AI plan and 1K without; Nano Banana Pro is "Available with a Google AI Plan," and once you hit the daily Nano Banana 2 quota you cannot redo images with Pro. Every image carries the visible watermark and SynthID. Google's overview page lists Nano Banana 2 at no charge with higher limits on paid plans, and says "limits apply."

Canva. Canva's AI photo tools live inside its design editor, so the staging prompt sits next to the listing flyer you are making. Canva's App Store listing shows the app free with in-app purchases, Canva Pro at $14.99 a month or $119.99 a year among the prices listed, and a 4.9 rating from 3.6 million ratings. Canva's own help pages for its AI editing tools and its AI product terms returned 403 to us, so we cannot quote usage limits or terms; read them before you rely on the tool for a listing.

Photoshop (Generative Fill). The closest thing to a professional tool among the four: select an area, "Write a prompt describing the object or scene you want to generate," and Photoshop fills the selection on its own layer, with the rest of the photo untouched. Uploads are JPEG or PNG up to 40 MB and 8,000 × 8,000 pixels, and a membership or free trial gives "a set number of Generative Credits." Because only the selection is regenerated, Generative Fill is the one option here that can leave the window and floor pixels alone, if your selection does. It still invents furniture at whatever scale it judges right, and Adobe's page says copyright ownership of the output "will depend on your local jurisdiction."

The prompts people use, and what to add to them

Search suggestions for "virtual staging prompt" are a sign that people are already staging listings this way. A typical prompt is short: "Add a modern gray sofa, a coffee table and a rug to this empty living room. Keep everything else exactly the same." The second sentence is doing all the work, and the models only partly obey it. Three additions help, none of them guarantees anything:

• Name the fixed features you want untouched. "Do not change the windows, the view through them, the floor, the ceiling, the moldings, the radiator or the doorway." A list is easier for the model to respect than "everything else."
• Describe scale. "A three-seat sofa about 220 cm wide, against the long wall, not blocking the window." Without it, furniture comes back at the size that looks best in the frame, which is often too small for the room.
• Ask for the same furniture again in the next angle. For the second and third photos of a room, upload the first staged result as a reference and say "the same sofa, table and rug, seen from the doorway." Google's API allows up to 14 reference images; consistency across angles is the thing the vendors themselves flag as hard.

Then check, every time, against the original: window frames and panes, the view, the floor pattern and its direction, doorways, wall color, built-ins, light fittings and the room's proportions. If anything has moved, do not use the image. A label does not cure a changed room.

General models vs purpose-built staging tools

Purpose-built staging tools are built on the same kind of model, but around the one job, with furniture libraries, room-type presets, a label stamped on the output and, in several cases, an explicit promise about the room. Virtual Staging AI's homepage says its AI "Leaves room layout untouched"; Shhots lists what it never changes (walls, windows, doors, floors, ceilings, built-ins, sightlines). Those are promises, not proof, and the same check against the original applies, but the tools are at least aimed at the problem. Prices from their public pages, checked October 20 and 21, 2026:

Read the first two rows against the rest. A general model costs nothing extra if you already pay for the chat, and for one hero image of an empty room, checked carefully, it can be enough. Once you need three angles of the same room, several rooms, or a label and an original filed with the MLS, the purpose-built tools are cheaper in time and safer in outcome. Our best virtual staging apps roundup and Virtual Staging AI alternatives compare them in more detail.

The disclosure duty is the same whatever made the picture

No MLS rule asks which model you used. CRMLS answers the question directly: "Yes, AI-altered images count as digitally altered images." NAR's Code of Ethics, Article 12, requires REALTORS® to "present a true picture" in their advertising, and Standard of Practice 12-10 bars "use of misleading images." In California, AB 723 has required since January 1, 2026 that any digitally altered image in material advertising a home for sale carries a statement that it was altered and a link, URL or QR code to the original, unaltered image; only lighting, sharpening, white balance, color correction, angle, straightening, cropping and exposure are exempt. Virtual staging is covered by name in the MLS guidance and plainly by the statute's "add, remove, or change elements" wording.

So a ChatGPT-staged room needs exactly what a BoxBrownie-staged room needs: a label on or next to the image ("Virtually staged"), the unaltered photo filed or linked where your MLS wants it, and no change to fixed features. The practical difference is that with a general model you also have to prove to yourself that the fixed features did not change, because nothing in the tool promised they would not. Compliance is already thin: in HousingWire's report on Coraly's review of just under 40,000 primary listing images, 10.8% showed signs of alteration and over 90% of those had no visible disclosure. Our explainers: California AB 723, AI photo editing and MLS rules and AI-generated listing photos.

Where October's labeled Empty and Staged rooms fit

October also uses an AI image model (OpenAI's GPT Image), but it starts from a guided iPhone capture of the room rather than a photo you upload, and it stages per room rather than per image. On any room captured with the October app (iPhone 12 or newer), you can order an Empty room ($19) with the furniture removed, or a Staged room ($29) in Scandinavian, Modern or Classic. Every delivered photo of that room comes in the chosen look, each image carries a visible "Virtually staged" or "Virtually emptied" label, and the model is told to keep walls, floors, windows, fittings, built-in furniture and the view as captured, with each photo checked against the capture and redrawn if the room changed. Pro plans ($399 a month, 6 places) include 3 such rooms a month. Every October photo, staged or not, is drawn by AI models, and no human review is promised, so the same look-before-you-publish rule applies.

The limit is the mirror image of the general models' strength: October only stages rooms you captured with its app, so it cannot touch the photos already on your listing. If those are what you have, use one of the staging tools above, or a general model with the checks on this page, and label the result. To decide whether a room should be shown staged or empty, see virtual staging vs empty room.

Step by step

1. Start from a good, level, unaltered photo

   Shoot the empty room in daylight with the phone level, and keep that original file. You will need it for the MLS and for the comparison at the end.

2. Write the prompt with the fixed features named

   Describe the furniture, its scale and placement, then list what must not change: windows, view, floor, ceiling, moldings, doors, built-ins, wall color.

3. Generate, then compare side by side

   Put the result next to the original at full size. Check every window pane, the view, the floor pattern, the doorways and the proportions. If anything moved, discard it.

4. Repeat for the other angles with the first result as reference

   Upload the approved image with the next photo and ask for the same furniture seen from the new spot. Check that angle the same way.

5. Label it and file the original

   Add "Virtually staged" on or next to each image, upload the unaltered photo where your MLS requires it, and in California add the link or QR code AB 723 asks for.

6. Decide if the next listing deserves a purpose-built tool

   If you stage more than one room or angle, a staging tool or a service with a designer saves the checking, and October stages captured rooms per room with the label built in.

Questions