Google released Nano Banana 2.1 on October 6 with four headline improvements. We tested each one on real property photos, against Nano Banana Pro and GPT Image 2.5 Sunburst, with the same photo and the same instruction for every model.
The short version: Nano Banana 2.1 is the most precise of the three at changing only the area you mark, and it climbed well past Nano Banana 2 and Nano Banana Pro on the public leaderboards. Nano Banana Pro still looks more like a real photograph and held the room better at 4K. GPT Image 2.5 Sunburst still leads every leaderboard and made the best-designed flyer.
What Google says improved
- Visual quality and realism at 1K, 2K and 4K.
- Mask-based editing: change a marked region and leave the rest alone.
- Subject consistency across repeated edits.
- Text rendering and infographic layout.
If you build on the Gemini API, one date matters more than any of that: Google's deprecations page lists Nano Banana 2 (gemini-3.1-flash-image) for shutdown on October 29, 2026, with gemini-nano-banana-2.1 as the replacement. Nano Banana Pro has no shutdown date.
The scores: Arena leaderboards
Arena ranks image models by blind human votes. Its original post includes the full Text-to-Image, Image Edit and Multi-Image Edit charts. The Multi-Image Edit chart is the one that places Nano Banana 2.1 fourth at 1,431 points:
Nano Banana 2.1 by @Google has landed in the Text-to-Image Arena, Image Edit Arena and Multi-Image Edit Arena! It landed: - #4 in Multi-Image Edit with 1431 pts - #5 in Text-to-Image with 1328 pts - #6 in Image Edit with 1428 pts This release is a substantial improvement from Show more
Multi-Image Edit
Text-to-Image
Image Edit
Meet Nano Banana 2.1, our latest image generation and editing model 🍌 This upgraded version outperforms our previous models across the board, with notable leaps in visual design, mask-based editing, and subject consistency to help you create more natural-looking images.
Select a chart in the post to zoom in without leaving this article. The original-post link opens X separately. These are Arena's published charts, separate from our room tests below.
Scroll sideways to compare all three leaderboards.
| Model / variant | Image Edit | Multi-Image Edit | Text-to-Image |
|---|---|---|---|
| GPT Image 2.5 Sunburst | 1,524 | 1,527 | 1,425 |
| GPT Image 2.5 Flare | 1,481 | 1,476 | 1,398 |
| Nano Banana 2.1 | 1,428 | 1,431 | 1,328 |
| Nano Banana 2 | 1,387 | 1,370 | 1,261 |
| Nano Banana Pro | 1,386 | 1,367 | Not shown |
| Nano Banana Pro 2K | 1,390 | 1,364 | Not shown |
| Grok Imagine Image 2.0 (canvas) | 1,439 | Not shown | 1,321 |
| Grok Imagine Image 2.0 (low) | 1,425 | Not shown | 1,297 |
| Grok Imagine Image Quality | 1,391 | Not shown | Not shown |
Source: the three charts in Arena's original post. Compare scores within the same column. These measure human preference, not staging accuracy, and are a dated snapshot rather than live rankings.
Not shown means the variant is absent from the published chart. Pro and Pro 2K are separate entries. Arena's named Grok variants are not a confirmed match for the Grok endpoint in our room tests.
In the published charts, Sunburst leads Nano Banana 2.1 by 96 points on both editing boards and 97 points on text to image. Nano Banana 2.1 is above both Pro entries on the editing boards. The text-to-image chart does not show Pro, so it cannot establish a direct comparison there. Arena's post quotes gains over Nano Banana 2 that differ from the gaps in its attached charts; the table above follows the chart values exactly.
Arena measures which picture people prefer overall. It does not measure whether the window in your listing photo stayed where it was. That is what our tests are for.
How we tested
Real property photos from our sample set, the same one-line instruction for every model, sent through the same pipeline our studio uses. One run per model unless noted, on October 7, 2026. One run can be lucky or unlucky, so treat each result as a sample, not a verdict.
Test 1: visual quality and realism
Instruction: "Stage this empty living room in warm modern for a listing."
At 2K, all three kept the two-pane window, the doorway, the ceiling light and the camera. All three staged the room convincingly. Nano Banana Pro's light and leather read most like a photograph; Nano Banana 2.1 is close behind and a little cleaner, closer to a polished render.
At 4K the picture changed.
Nano Banana Pro held the room at 4K. Nano Banana 2.1 pulled the camera back and hung curtains over the window, which is exactly the kind of change you cannot have in a listing photo. Up close, both are sharp; Pro's textures, like the swirl in the ceramic bowl and the dried leaves, look more like a real photograph.
Verdict: the quality claim holds at 1K and 2K. At 4K, Pro was the more reliable of the two in our run.
Test 2: mask-based editing, or "change only what I marked"
First, something Google's announcement does not spell out: the Gemini API has no mask parameter. Its image guide describes editing with text instructions and reference images only. The mask editing Google talks about lives in its own apps. Through the API, the way to say "change this area" is to send the photo plus a second, marked-up copy of it, which is what our studio does when you mark an area.
So we marked only the back run of benchtop, along the sink wall, and asked:
Instruction: "Change the marked benchtop section to honed white marble. Leave the side benchtops, cabinets, sink, tap, tiles and everything else exactly as they are."
Close up on the benchtop, where the differences are:
- Nano Banana 2.1: changed the marked section only. Both side benchtops stayed granite.
- Nano Banana Pro: changed the marked section and carried the marble onto the right-hand benchtop beside the stove.
- GPT Image 2.5 Sunburst: changed the marked section only.
- Sunburst with OpenAI's native mask: our marked area covered the sink, and the model treated the sink as part of the benchtop: it replaced it with marble and left the tap standing on a bare slab. A native mask means "anything inside may be redrawn", so the mark has to leave out what must survive.
Verdict: the claim holds. Nano Banana 2.1 was the most precise Google model here, and it beat Nano Banana Pro.
Test 3: subject consistency across edits
We took each model's own staged room and asked for a second edit.
Instruction: "Change only the sofa to deep green velvet. Keep everything else in the room exactly as it is."
All three changed the sofa and nothing else: the rug, the table, the lamp, the plants, the artwork, the window and the camera stayed put.
Verdict: the claim holds for Nano Banana 2.1, but this is no longer what separates the three. They all passed.
Test 4: text and layout
Instruction one: add a real estate sign in the lawn reading "JUST LISTED" and "3 BED · 2 BATH", and keep the house, lawn and driveway as they are.
All three spelled the sign perfectly. The difference was everything around it: Nano Banana 2.1 turned the lawn bright green and planted a new garden bed it was told to leave alone. Nano Banana Pro left the house and lawn untouched. Sunburst added a few flowers.
Instruction two: turn the photo into a listing flyer with a headline, an address, a price and three labelled icons.
Every word and number was right in all three. Sunburst produced the most finished layout. Nano Banana 2.1 redesigned the flyer as a clean portrait page. Nano Banana Pro's text panel sat over the lawn and felt crowded.
Verdict: text rendering is solid on all three now. Layout taste still favours Sunburst.
The scorecard
Scroll the table sideways to compare all three models.
| Google's claim | Nano Banana 2.1 | Nano Banana Pro | GPT Image 2.5 Sunburst |
|---|---|---|---|
| Visual quality at 2K | Kept the room, polished | Kept the room, most photographic | Kept the room |
| Visual quality at 4K | Moved the camera, added curtains | Kept the room | Not offered at 4K |
| Change only the marked area | Precise | Spilled onto a second benchtop | Precise |
| Consistency across edits | Passed | Passed | Passed |
| Text on a sign | Correct, but edited the lawn | Correct, scene untouched | Correct, added flowers |
| Flyer layout | Clean portrait design | Crowded | Most finished |
Which one to use
- Everyday staging and precise fixes on listing photos: Nano Banana 2.1. It changed exactly what was marked and stayed faithful at 1K and 2K.
- The most photographic result, or a 4K print: Nano Banana Pro.
- Flyers, layouts and the overall look people prefer: GPT Image 2.5 Sunburst.
Whichever you use, check the fixed things before you publish: window count, door positions, the line where the wall meets the ceiling. And if the photo will sell or rent the property, say it was virtually staged. Our FAQ has the short version.
Try them on your own photo
Nano Banana 2.1 joins Nano Banana Pro and GPT Image 2.5 in our studio, where Nano Banana 2.1 uses far fewer credits per render than Nano Banana Pro. Mark the area you want changed, pick a model, and the studio shows the credits before you run it, so you can compare models on your own room. The first five renders are free with a watermark; plans start at $15 a month. Try virtual staging on your photo or see the pricing page.
Tests run on October 7, 2026: one run per model per test on our sample photos, except Nano Banana 2.1's 2K staging, which we ran twice with the same result. Arena scores from arena.ai on October 6, 2026. Results vary between runs and photos. Nano Banana is a Google product and GPT Image 2.5 an OpenAI product; Astravisualize is independent and not affiliated with either.
- nano-banana-2-1
- nano-banana-pro
- gpt-image-2-5
- image-models
- virtual-staging
