How an Agent in Off The MRKT Tested AI Models for Staging Concepts
Published November 21, 2026 · Source: Off The MRKT
An unnamed agent writing in Off The MRKT tested four Nano Banana image models on an empty living room to make virtual staging concepts for owners. They say 14 images cost $0.79 and are good enough for owner conversations, not listing photos.
Result, in their words: Total spend was $0.79 for 14 images. The author says the models are "good enough for owner conversations and direction boards, and not yet for listing-grade photos I'd sign my name to."
From the original
Total spend was $0.79 for 14 images. Today these models are good enough for owner conversations and direction boards, and not yet for listing-grade photos I'd sign my name to. Lighting is a bit too even, shadows a bit too soft, and furniture placement drifts.
What they did
An agent who writes in Off The MRKT, and who says they do light staging for their own listings, tested whether AI image models could produce staging concepts to show an owner before any furniture is moved. The article does not name the author. They did not use a client's apartment. Instead they generated a stand-in empty pre-war living room with Nano Banana Pro and a furniture board with Nano Banana 2 Lite.
How it worked
The author says they gave the empty room and the furniture board to four models on APIMart: Nano Banana Pro, Nano Banana 2, Nano Banana 2.1 and Nano Banana 2 Lite. Each got the same instruction, with exact furniture positions and the line "Do not move, add or remove any architectural element. Only add furniture and decor."
They then compared each result with the empty room pixel by pixel in five zones (each window, the radiator, the doorway and the ceiling line) and checked each zone by eye.
They ran three tests:
- Staging from the board. By their account, all four models kept the architecture. Only Pro and 2 Lite put the sideboard where asked. The other two moved it to a side wall, over an outlet.
- Text-only style variants. Nothing was redrawn, but features got hidden. Nano Banana 2 put the sofa in front of the radiator, and Nano Banana 2.1 put plants across part of a window and the doorway.
- A one-item revision. Swapping the green sofa for a brown leather one was "the best result of the day". Outside the sofa, they report under 1% of pixels changed in all four models.
The result
The author says the total spend was $0.79 for 14 images. Their verdict is that the models are good enough for owner conversations and direction boards, but not for listing-grade photos. They say the lighting is too even, the shadows are too soft, and furniture placement drifts.
Their own rules for any staged image are a "virtually staged" label on the image itself, the real empty photo next to it, and no structural changes or hidden features.
How to try it on your listings
This is our advice, not the author's:
- Use AI staging first as a concept tool. Agree a furniture direction with the owner before you spend on physical staging.
- Write exact positions into your prompt and tell the model to leave the architecture alone. Then check the windows, radiators, outlets, doors and views against the original photo.
- Test with a small change, such as swapping one item, before you rerun a whole room.
- Virtual staging changes how the property looks. It may need a disclosure label, and your MLS or brokerage may restrict or prohibit it, so check their rules before you post. Keep the unedited photo alongside.