Can ChatGPT Do Virtual Staging? We Tested It on the Same Room
We ran one empty room through ChatGPT’s image model and three purpose-built staging models. ChatGPT produced a beautiful room — just not the same room. Here are the outputs, side by side.
“Can ChatGPT do virtual staging?” turns up in Google’s People Also Ask box under nearly every virtual staging search — the head term, the software comparisons, the pricing queries, all of them. It is the single most repeated question in the category and the answers are uniformly vague.
ChatGPT Virtual Staging Test Results
What aspect ratio did the test room have, and what aspect ratio did ChatGPT return it at?
Answer: 1.286:1 input, 1.500:1 outputThe article states the test room measured 1152 × 896 pixels with an aspect ratio of 1.286, and that ChatGPT forced it to 1.500:1. This is a structural limitation of OpenAI's image endpoint, which emits only fixed sizes.According to the test results table, which model was both the cheapest and fastest?
Answer: Gemini 3.1 Flash LiteThe table shows Gemini 3.1 Flash Lite cost $0.0339 per image and took 6.3 seconds, making it the lowest cost and fastest of all models tested.What specific architectural change did gpt-image-1-mini make to the windows?
Answer: It widened the two separate windows into what reads as a single window unitThe article explicitly states that gpt-image-1-mini was worse than gpt-image-1, with the two separate windows widening into what reads as a single window unit, and the radiator vanishing completely.
So we tested it properly. One empty room, one prompt, five models, outputs compared at full size.
The short answer: ChatGPT will give you a beautiful staged room. It will not reliably give you your staged room. And for a listing photo, that distinction is the entire job.
The test
The room is a generated photorealistic empty living room — deliberately synthetic so it is licence-clean and reproducible, and so the geometry is unambiguous. Two rectangular windows on the left wall, a radiator beneath them, a doorway on the right, oak floor, crown moulding. It measures 1152 × 896 pixels, an aspect ratio of 1.286.
Every model got the identical prompt: stage as a warm modern living room, add sofa, armchair, coffee table, rug, floor lamp, plant and wall art, and do not change the walls, ceiling, windows, window frames, doorway, radiator, baseboards, floor material or camera angle.
That last clause is not decoration. Every MLS rule on virtual staging says a version of the same thing: you may add furniture, you may not change the building.
The results
| Model | Cost per image | Time | Aspect kept | Architecture kept |
|---|---|---|---|---|
| Gemini 3.1 Flash Lite | $0.0339 | 6.3s | ✅ 1.286 | ✅ |
| Gemini 2.5 Flash Image | $0.0391 | ~9s | ✅ 1.286 | ✅ |
| Gemini 3.1 Flash Image | $0.0678 | 11.6s | ✅ | ✅ |
| GPT-image-1 (ChatGPT) | — | — | ❌ forced 1.500 | ❌ |
| GPT-image-1-mini | — | — | ❌ forced 1.500 | ❌ |
Costs are the actual billed amounts returned by the API, not list-price estimates.
What ChatGPT got wrong
It cannot return your photo’s shape. This is structural, not a prompting failure. OpenAI’s image endpoint emits fixed sizes — 1024×1024, 1536×1024, 1024×1536. Our 1.286:1 room came back at 1.500:1. Before you evaluate a single piece of furniture, the room has been re-proportioned and the framing changed. No prompt fixes that, because the constraint is the API, not the model’s understanding.
It redrew the windows. On close crop, gpt-image-1 returned windows that are shorter, wider and differently spaced than the originals, with the wall-to-ceiling junction moved. gpt-image-1-mini was worse: the two separate windows widened into what reads as a single window unit, and the radiator vanished completely.
Both outputs are genuinely attractive rooms. Shown on their own, most people would call them excellent staging. Put them next to the original and they are a different property.
What the purpose-built models got right
All three Gemini image models returned the room with the same two windows, the same mullion grid, the same spacing, the same sill and trim, the same doorway, the same floor, and the same aspect ratio. The furniture is added; the building is untouched.
One honest caveat about the tool we built on this: in one run, the model placed a sideboard where the radiator was, hiding it. Real stagers put furniture in front of radiators all the time, so this is defensible — but it is a change to what the buyer can see, and worth knowing about.
Why this is a compliance problem, not a taste problem
If this were only about quality, the answer would be “use a better tool”. It is not.
NorthstarMLS prohibits virtual staging that alters “walls, floors, doors, windows, roofing, siding, ceilings, driveways or site grading.” Canopy MLS prohibits adding features that do not exist. CRMLS, enforcing California’s AB 723 since 1 January 2026, prohibits altering “walls, flooring, landscaping or dimensions.”
An image with re-proportioned windows violates those rules on its face — and the agent has no idea, because they disclosed the thing they did (added furniture) and not the thing the model did (moved the window). Disclosure does not cure it. You cannot label your way out of depicting a property that does not exist.
That is the real answer to “can ChatGPT do virtual staging”. Not the quality is a bit worse. It is: the failure mode is invisible to you and puts you the wrong side of an MLS rule.
When ChatGPT is genuinely useful here
None of this means the tool is useless for listing work. It means it is the wrong tool for this specific job. Things it does well:
- Writing the listing copy around the staged photo. Genuinely good, and we have twelve prompts for it.
- Choosing a furniture style for a property. Describe the house and ask what would suit the likely buyer — it is a reasonable design consultant.
- Concept renders that are not listing photos. Renovation ideas for a seller conversation, mood boards, “what could this room become”. Nobody is treating those as a record of the property.
- Drafting your disclosure language.
The line is whether the output will be presented as a photograph of the property. Inside that line, use a model that preserves geometry. Outside it, ChatGPT is fine.
What to use instead
Purpose-built staging tools use image-editing models with the geometry constraint built in, and they return your photo at your photo’s dimensions. That includes the paid services — ApplyDesign, VirtualStagingAI — and our own free virtual staging tool, which runs the cheapest model in the table above at three photos a day, free, with the MLS disclosure label applied by default.
Whatever you use, do the check that took us five minutes: open the original and the staged version at full size, side by side, and look at the windows. It is the fastest way to catch a model that has quietly redecorated the architecture.
FAQ
Can ChatGPT do virtual staging?
It can generate a staged room from your photo, but in our testing it did not preserve the room. OpenAI’s image endpoint only returns fixed aspect ratios, so a listing photo comes back re-proportioned, and both GPT image models we tested altered the windows. For a photo that will represent an actual property on the MLS, use a purpose-built staging tool.
Why does ChatGPT change the room?
Two reasons. Its image endpoint emits only fixed sizes (1024×1024, 1536×1024, 1024×1536), so any photo that is not one of those shapes is re-framed. And it is a general-purpose generative model optimised for producing a good image, not for reproducing the input image with additions — so it treats architecture as something it can improve.
Is it against MLS rules to stage with ChatGPT?
Staging with any tool is fine if you disclose it. The problem is that MLS rules — Canopy, NorthstarMLS and CRMLS among them — separately prohibit altering walls, windows, doors, floors or dimensions. If the model moves a window, the image breaks the rule regardless of your disclosure.
Which AI is best for virtual staging?
In our test the Gemini image models preserved the room’s architecture and aspect ratio; the GPT image models did not. Gemini 3.1 Flash Lite was both the cheapest at $0.0339 an image and the fastest at 6.3 seconds. Most purpose-built staging products are built on models of that type.
How much does AI virtual staging cost per photo?
The raw model cost is a few cents — we measured $0.0339 to $0.0678 an image. What vendors charge is a different question, driven by monthly photo caps rather than per-image cost. VirtualStagingAI starts at $16 a month for six photos; ApplyDesign’s auto mode works out at $10.50–$15 an image. Ours is free for three a day.
Can I just ask ChatGPT to keep the windows the same?
We did. The prompt explicitly listed the windows, frames, doorway, radiator, baseboards, floor and camera angle as things not to change. The windows changed anyway, and the aspect ratio changed because the API cannot do otherwise.