ChatGPT vs Gemini for virtual staging — the regenerate-vs-inpaint distinction
Same empty dining room. Same furniture list. Same 60 seconds. Two different rooms come back — and one will get a buyer's agent flagging the listing.
The Hendersonville agent who staged her Old Hickory Lake ranch in ChatGPT and the one who staged it in Gemini aren't running the same workflow. They think they are. They aren't. The model under the hood is different, and the difference shows up the day a buyer walks the property.
The detail nobody talks about
ChatGPT regenerates the room. Gemini 2.5 Flash inpaints it. Those two verbs are the whole game.
Upload a phone photo of an empty dining room and ask ChatGPT to add a table and chairs — the model produces a new image inspired by yours. Walls might be a few inches longer. The window might be wider. The ceiling fan might quietly become a chandelier. (Simon Willison's Mar 25, 2025 walkthrough of 4o image generation is the cleanest technical write-up — the model produces a fresh image conditioned on yours, it doesn't edit the pixels you uploaded.)
Upload the same photo to Gemini 2.5 Flash with the right prompt and the model edits the photo. Walls stay put. The window stays sized. The fan stays a fan. Only the furniture changes.
Nathan Cool — a real-estate photographer with 15 years on the camera — warned about this in Fstoppers (Alex Cooke, Apr 7, 2026): "Many AI HDR tools aren't actually editing your photos. They're remaking them. That distinction matters legally." He documented AI quietly altering wall lengths and window sizes on real listings. Naumann Legal (Mar 2026) lays out the civil-fraud claim that follows when a regenerated photo hides a real construction defect.
That's regenerate in production. The buyer's agent walks the Compass listing in Brentwood, opens Zillow on the porch, sees a room that doesn't quite match. The deal doesn't die on the photo — it dies on the trust loss that follows.
Inpainting — editing pixels in a defined region while preserving everything outside it — is what Gemini does on the right prompt. The move for any photo that ships to MLS.
Side-by-side: the actual difference
Table: ChatGPT vs Gemini for virtual staging — verified 2026.
| Dimension |
ChatGPT (GPT-4o image gen) |
Gemini 2.5 Flash Image |
| Underlying mechanism |
Regenerate — produces a new image inspired by yours |
Inpaint — edits the actual photo, preserves structure outside the edit region |
| Prompt that locks the mode |
"Stage this empty dining room in [style]. Match the original lighting." |
"Using the provided image, change only the [furniture]. Keep everything else exactly the same — walls, windows, floor, lighting, composition." |
| Output behavior |
Walls may lengthen. Windows may widen. Fixtures may swap. The room you get is like yours, not yours. |
Walls stay put. Windows stay sized. The room you get is yours, with the new furniture dropped in. |
| Speed |
~60 seconds per image |
~60 seconds per image |
| Cost |
$20/mo (ChatGPT Plus, unlimited) |
$20/mo (Gemini Advanced, unlimited) |
| Best use case |
Pre-vis, concept work, "what could this empty room look like" for the seller |
Listing photos, MLS uploads, anything that has to match the physical property |
| Failure mode under pressure |
Hallucinates structural detail when prompt is loose |
Refuses or no-ops when the edit region is ambiguous — safer failure |
| Disclosure overlay built in |
No |
No (both require manual overlay step) |
Price tied. Speed tied. Mechanism is the whole separation.
ChatGPT isn't broken. Regenerate is the right tool for the right job — pre-vis for a seller meeting, concept work where dimensions don't have to be honest. The model is creative on purpose. The problem is using the regenerate model on a listing photo. That's a category error.
Sources: Google Developers Blog, How to prompt Gemini 2.5 Flash Image generation for the best results, Aug 28, 2025. Fstoppers, The legal risk most real estate photographers don't know they're taking, Alex Cooke, Apr 7, 2026. Simon Willison, Introducing 4o image generation, Mar 25, 2025 on the regenerate mechanism.
The Gemini inpaint prompt template — copy-paste ready
The Google Developers Blog (Aug 28, 2025) shipped the template that locks Gemini 2.5 Flash Image into inpaint mode. Verbatim:
"Using the provided image, change only the [specific element]. Keep everything else in the image exactly the same, preserving the original style, lighting, and composition."
That sentence is the whole game. Without it, the model drifts toward regenerate. With it, the model holds structure.
Working version for a Hendersonville dining-room stage. Copy-paste into Gemini 2.5 Flash Image with the photo attached:
Using the provided image of an empty dining room, change only the
following — add:
- six-seat round wood dining table
- four upholstered linen chairs
- a simple linen runner
- a clear glass vase with eucalyptus
- a wool area rug under the table
Keep everything else in the image exactly the same, preserving the
original style, lighting, and composition. Do not change the wall
color, the hardwood floor, the windows, the ceiling fan, or the
natural afternoon light. Do not change room dimensions or add
architectural features.
That's the workflow Sarah at Compass Hendersonville is running on her Old Hickory Lake ranch. Same template, different room, different furniture list. The "keep everything else in the image exactly the same" line is doing the load-bearing work.
Notes on the template:
- "Change only the" matters. Drop it and Gemini gets permissive.
- List furniture explicitly. Vague prompts drift toward regenerate.
- "Preserving style, lighting, composition" is non-negotiable — verbatim Google language.
- Add the negative-instruction line. "Do not change room dimensions or add architectural features." Belt and suspenders. Karpathy's "Software 3.0" framing reads here too — natural language is the program — the precision of the constraint sentence is the precision of the output.
Source: Google Developers Blog, Aug 28, 2025.
The ChatGPT prompt — when regenerate is what you want
Real cases for regenerate exist. Pre-vis is the big one.
Sarah's seller has an empty primary suite. Sarah wants to show the seller what the room could look like — a vision image to convince the owner to pay for a physical stager. Not for MLS. For the seller's email.
Regenerate wins here. ChatGPT produces a more aspirational image — softer light, idealized angles, a room that looks better than the bones suggest. That's the point. You're selling a vision, not documenting a property.
Copy-paste into ChatGPT with the photo attached:
Using the attached photo of an empty primary bedroom as inspiration,
generate a fully designed concept image showing the room staged in
a transitional style. Include:
- a king bed with a linen upholstered headboard
- two warm-wood nightstands with brass lamps
- a textured area rug
- soft afternoon light from the windows
- one piece of subtle wall art above the bed
This is a concept image for a seller meeting, not a listing photo.
Lean toward an aspirational, magazine-style result. Match the general
proportions of the room shown in the photo.
Two things to notice. "Generate a concept image" gives the model permission to regenerate. "Match the general proportions" tells the model the bones are a guide, not a constraint.
This output is for a seller email — not for MLS. Put a regenerated concept image on a listing without disclosure and a paired original and you're outside NAR Code of Ethics Article 12 — the "true picture" standard.
The line: regenerate for the seller meeting, inpaint for the listing.
Side-by-side photo comparison — described, not embedded
Picture two staged photos of the same Hendersonville dining room. Sarah ran the inpaint prompt in Gemini at 9:14 AM and the regenerate prompt in ChatGPT at 9:16 AM. Same source photo, two minutes apart.
The Gemini output. Six-seat round wood table, four linen chairs, glass vase with eucalyptus, wool rug. Wall is the same off-white the seller painted in 2019. Lake-facing window is the same width, same height. Ceiling fan is still there — the brass four-blade that came with the house. Hardwood floor has the same scratches near the back door. Afternoon light comes in at the same angle.
The ChatGPT output. Same furniture list. Wall is almost the same off-white — slightly warmer, a touch of cream the original didn't have. Window is almost the same — maybe two inches wider, the framing slightly cleaner. Ceiling fan is gone, replaced by a three-bulb chandelier. Hardwood floor is glossier, no scratches. Afternoon light a tone warmer.
A casual scroll won't catch it. A buyer's agent walking the property at 4:30 PM on a Saturday will catch every one. The buyer notices the buyer's agent noticing. Trust drops a notch. The offer comes in $8K under.
That's the cost of running regenerate where inpaint was the job.
Gemini with the disclosure overlay baked in is the operator move. ChatGPT on MLS without disclosure is the failure mode the staging association called out — "bait and switch … and when the buyer gets there it's a let down" (Real Estate Staging Association via HousingWire, 2025).
Phone-First Staging — the pattern, in one paragraph
The whole frame is on the pillar page. Short version — below $1M list price, your phone plus the foundation model you already pay for stages a room in 60 seconds for $20/mo. You need the inpaint prompt above, the disclosure overlay, the unaltered original next to the staged photo, and a one-line note in the listing description. Above $1M, the three thresholds shift — hire the photographer at $1M, hire both photographer and stager above $1.5M.
Jason Haber, the NYC agent, in PetaPixel (Matt Growcoot, Oct 27, 2025): "Why would I send my photos to a virtual stager when I can just do it in ChatGPT for free in 45 seconds?" Right on the price math. More right if he'd specified Gemini inpaint for the listing photos. That's the small upgrade that takes Phone-First Staging from "it works" to "it ships."
FAQ
Is ChatGPT or Gemini better for virtual staging?
Neither, in isolation. Gemini 2.5 Flash Image with the inpaint prompt is better for listing photos — preserves structure, walls, windows, lighting. ChatGPT is better for pre-vis and concept work — regenerate is creative on purpose, which is what you want when selling a vision to a seller. Both run $20/mo. Both take ~60 seconds. The choice is about the use case.
What's the exact Gemini prompt for inpaint mode?
Verbatim from the Google Developers Blog (Aug 28, 2025): "Using the provided image, change only the [specific element]. Keep everything else in the image exactly the same, preserving the original style, lighting, and composition." Drop either line and the model drifts toward regenerate. Both are load-bearing.
Why does ChatGPT lengthen walls and widen windows?
The model isn't editing your photo. It's generating a new image inspired by it. Simon Willison's coverage of 4o image generation (Mar 25, 2025) walks the technical mechanism. Without the inpaint instruction set, the model quietly tunes proportions — longer walls, wider windows, swapped fixtures, cleaner floors. Nathan Cool documented this in Fstoppers (Apr 7, 2026). On a $539K Hendersonville ranch the gap may be small. On a Brentwood listing where buyer's agents tour back-to-back, it's a deal-killer.
Do I still need to disclose if I use Gemini inpaint?
Yes. Inpaint is structurally honest, but the photo is still digitally altered. NAR Code of Ethics Article 12 applies. You still need the "Virtually Staged" overlay at 14-18pt minimum (Bay East MLS standard 2025), the unaltered original in the listing sequence, and a one-line description note. The model choice doesn't replace the disclosure step. It just makes the photo defensible.
Can I use both ChatGPT and Gemini on the same listing?
Yes, for different jobs. Gemini inpaint for the MLS photos. ChatGPT regenerate for the seller meeting concept image. Two tools, two jobs, $40/mo total — both already on the operator's bill.
Which one handles batches better?
Neither. Both foundation models drift in style consistency across 50+ image batches. That's where the paid tools earn their fee. For 5-10 photos on a sub-$1M listing, the foundation model wins. For a 50-photo Brentwood listing with a 360° tour, the paid tool starts to pay for itself.
The closer
"ChatGPT or Gemini for virtual staging" is the wrong question. The right question — "what mode am I in, and does my prompt lock it there?"
Inpaint for the listing. Regenerate for the seller meeting. The prompt template above does the work. The disclosure overlay does the rest.
If you're shipping enough listings that the workflow is starting to leak — disclosure skipped on rush jobs, prompt drift across rooms, seller meetings without pre-vis the night before — the question stops being "which model" and becomes "how do I run the whole pipeline." That's what The Listing Machine operationalizes. Four-week cohort, prompt stack tuned to your voice, the AI-Enhanced Realtor credential at the end. We work against your real listings — Old Hickory Lake, Cool Springs, Brentwood — not a hypothetical.
For everyone running one or two listings a month — the Gemini prompt is the move. The ChatGPT prompt is the move for the seller meeting. The disclosure step is the floor under both. Bookmark them. Run them. The room changes. The template doesn't.
Sources
Practitioners (real-estate agents + photographers):
Builder/technical creators:
Regulatory / disclosure (primary):
Last updated 2026-04-29.