Virtual Staging 12 Empty Rooms With an Image API: Cost by Model

The short answer

Staging 12 empty-room photos with an image edit costs $0.6675 on gpt-image-2.5 at medium quality and 2K, $1.80 on Nano Banana 2.1 at 2K, and $2.6700 at gpt-image-2.5 high quality and 4K. Those are Sume list prices per image from the public catalog (read 2026-10-08), one edit per room, no retries.

How an edit request looks

Use POST /v1/images or the older POST /v1/image-1.0/generate. Send prompt and image_urls with one to ten public HTTPS photos. For a masked edit, add mask_image_url so the walls and floor stay fixed and only the masked area changes. Docs say quality is low, medium or high, and that higher values suit finals and dense text.

Google's pricing page lists Nano Banana 2.1 (gemini-nano-banana-2.1) as a current model (read 2026-10-09). Sume lists nano-banana-2.1 separately and bills its own catalog price, so use the Sume figures below for budgeting here.

A sensible order

Run all 12 rooms at low quality first. At $0.2970 the whole set is a layout check: does the sofa sit where a sofa would, is the window light kept. Re-run only the rooms you keep at medium. If three rooms need a retry at medium that adds $0.1669.

Keep the original photo as the first image_urls entry. The num_images field takes 1 to 4, so asking for two style options per room doubles the line, not the request count.

• Mark staged photos as staged in the listing. Many markets expect that disclosure; check your local rules.
• Do not stage over damage or change fixed features such as windows or room size.
• Store the Sume media.sume.com URL from result.artifacts[], not a provider URL.

Next step

A staged still can also feed a short clip. See the 8-photo listing reel for the video cost.

What to check on each result

Compare each staged image with the source at full size. Look for furniture that blocks a door, rugs that bend at a wall, and shadows that point the wrong way for the window. A masked edit lowers these risks because the room shell is outside the mask, but it does not remove them.

The job result lists the image under result.artifacts[]. Store that Sume URL with the listing record, and keep the unstaged original beside it so you can show both on request.

Treat the 12-room figure as a floor, not a quote. A realistic plan adds one retry for every third room, and a retry costs the same as the first try. At gpt-image-2.5 medium that is four more images, or $0.2225. Set the budget with that in mind, and use max_spend_usd or dry_run on MCP calls so a loop cannot run past it.