TL;DR
“Nano Banana” is the nickname for Google’s image-editing models inside Gemini — not a separate app. As of June 2026 there are three: Nano Banana (gemini-2.5-flash-image, the original fast/cheap editor), Nano Banana 2 (gemini-3.1-flash-image, speed-optimized successor), and Nano Banana Pro (gemini-3-pro-image, the high-fidelity flagship that went GA in June 2026 with 4K output and the best in-image text). You reach all of them by uploading an image to a Gemini chat and describing the change in plain language. The skill is writing edits that name what to keep, iterating one change per turn, and saving intermediates so you can roll back. This guide covers which model to pick, the exact workflow, real pricing and limits, and the six edits that reliably go wrong.
What “Nano Banana” actually is
Nano Banana started as an anonymous entry on lmarena.ai that beat every other editor in blind tests in 2025. Google confirmed it was theirs — Gemini 2.5 Flash Image — and the nickname stuck so hard that Google now uses “Nano Banana” in its own marketing. There is no “Nano Banana” app or button. When you upload a photo to a Gemini chat and say “change the jacket to red, keep the face,” the model handling that request is one of the Nano Banana family.
By June 2026 the family has three members, and picking the right one matters more than any prompt trick:
| Model | Marketing name | Best for | Resolution | Reference images |
|---|---|---|---|---|
gemini-2.5-flash-image | Nano Banana | High-volume, low-latency edits | up to 2K | small sets |
gemini-3.1-flash-image | Nano Banana 2 | Fast everyday edits, more aspect ratios | 0.5K–4K | up to 10 objects + 4 characters |
gemini-3-pro-image | Nano Banana Pro | Hardest edits, legible text, 4K assets | 1K / 2K / 4K | up to 14 objects + 5 people |
(Model names and limits per Google’s Gemini API image docs, June 2026.)
The short version: in the Gemini app the model is chosen for you based on your plan and prompt complexity. If you call the API, pick Nano Banana Pro when you need readable text inside the image, 4K output, or up to five consistent people in one scene; pick Nano Banana 2 for fast, cheap iteration.
What it does well — and where it breaks
Strong:
- Targeted local edits that leave the rest untouched — swap a shirt color, remove a sticker, change the background — without touching a mask tool.
- Multi-turn iteration in one chat: “now make the sky orange,” “now zoom on the face.” The conversation remembers prior edits, so a 10-step chain stays coherent.
- In-image text (Nano Banana Pro): legible headlines, paragraphs, even multilingual text and translation, with single-line error rates Google reports as mostly under 10%. This is the single biggest jump over the original model, where text reliably warped.
- Identity preservation: the same person or pet across scenes, outfits, and angles. Pro holds up to five people consistent; you can blend up to 14 reference objects in one composition.
- Lighting, focus, and camera moves: day-to-night, shallow depth of field / bokeh, and angle or shot-type changes are now first-class controls in Pro.
Weak:
- Pixel precision. There are no millimeter-accurate selection brushes, layer masks, or exact color-value inputs. For geometric or pixel-exact work, finish in Photoshop or Affinity Photo.
- Heavy multi-subject scenes still lose secondary detail, and aggressive style swaps can break composition.
- Free-tier output is capped at 1K and carries a visible Gemini “sparkle” watermark plus the invisible SynthID signature. Paid and API outputs keep only the invisible SynthID — no visible branding.
Who this is for
Anyone with a base image to tweak — product shots, portraits, social posts, mock-ups — who wants natural-language edits instead of opening a desktop editor and building a mask.
Pricing and limits (June 2026)
In the Gemini app, image editing rides your Google AI plan rather than a separate fee:
| Plan | Price/mo | Nano Banana Pro images/day | Max resolution | Watermark |
|---|---|---|---|---|
| Free | $0 | tight daily cap (~a few) | 1K | visible sparkle + SynthID |
| Google AI Plus | $9.99 | up to 50 | up to 2K | SynthID only |
| Google AI Pro | $19.99 | up to 100 | up to 4K | SynthID only |
| Google AI Ultra | $99.99 | up to 1,000 | 4K | SynthID only |
(Gemini app daily caps per Google’s Gemini Apps limits help page, June 2026; “Google AI Pro” is the plan formerly called Gemini Advanced.)
If you call the API instead, Nano Banana Pro bills per generated image by resolution — roughly $0.039 at 1K, $0.134 at 2K, and $0.24 at 4K as of June 2026. The lower-cost Flash models (Nano Banana / Nano Banana 2) are the better choice for high-volume iteration; reach for Pro on the final, hero asset.
Step by step
- Go to
gemini.google.comand start a new chat (sign in with the Google account on the right plan — your plan also sets which model and resolution you get). - Upload the base image with the attach button. To preserve a person or product across edits, attach the clean reference shots up front rather than mid-chain.
- Describe only this turn’s change in plain language, and name what to keep:
Replace the background with a soft blue gradient. Keep the subject, pose, and lighting exactly the same. - Iterate one change per turn —
now make the sky orange, thennow zoom on the face— so the model builds on the previous result instead of averaging several edits. - Download each version you might want to revert to. The chat keeps history, but a fresh download is the only reliable way to lock in an intermediate at full resolution.
A worked example
Say you have product-shot.jpg of a bottle with a price sticker. A clean three-turn chain:
Turn 1: Here is product-shot.jpg. Remove the price sticker on the
bottle. Keep the label, glass reflections, and shadow untouched.
Turn 2: Now place the bottle on a marble surface with soft window
light from the left. Keep the bottle and label identical.
Turn 3: Now give me a 4:5 crop for Instagram at the highest
resolution available.Each turn changes one thing and states what to hold constant — that is the whole discipline.
Nano Banana vs other editors
Photoshop generative fill
Photoshop lives inside a full layer-and-mask editor, so it wins when you need millimeter-precise selection, non-destructive layers, or exact hex color values. Nano Banana is faster and friendlier for whole-image, natural-language edits — “make this look like dusk, keep the model’s face” — with no masking. Rule of thumb: geometric or pixel-exact → Photoshop; descriptive → Nano Banana.
Flux Kontext (Black Forest Labs)
Flux Kontext is another instruction-driven editor, often strong on hard composite edits where an inserted object must match shadow and perspective. It is slower and lives behind its own API or third-party UIs. Nano Banana wins on speed and conversational multi-turn iteration inside one Gemini chat; switch to Flux Kontext when a single composite keeps failing.
Seedream-edit (ByteDance)
Seedream-edit is notably strong on Chinese-language prompts and identity preservation for Asian faces. Nano Banana is more universal and integrated into the rest of Gemini. If your edits are heavily Chinese-prompted and portrait-focused, keep Seedream-edit as a second option.
Six mistakes that ruin edits
- Asking for many changes at once. The model averages them and you lose control. One change per turn.
- Not saying what to keep. Without “keep the subject, lighting, and composition,” the model is free to redraw them. State the keep-list every time.
- Forgetting the chat compounds. Your fifth message builds on edits 1–4. To return to the original, start a fresh chat and re-upload the original — you cannot un-edit backward in place.
- Expecting long in-image text on the Flash models. Original Nano Banana warps logos, signs, and paragraphs. Use Nano Banana Pro for legible text, or add the text layer in a real editor afterward.
- Using it for millimeter layout. Exact crops, alignment, and hex values belong in Photoshop on the final pass.
- Not saving intermediates. If iteration 7 is worse than iteration 4, you want iteration 4 still on disk. Download as you go.
FAQ
Q: Is Nano Banana a separate app from Gemini?
A: No. It is Google’s nickname for its Gemini image-editing models (gemini-2.5-flash-image, gemini-3.1-flash-image, and gemini-3-pro-image). You use them by uploading an image to a normal Gemini chat — there is no separate UI.
Q: What is the difference between Nano Banana and Nano Banana Pro? A: “Nano Banana” usually means the original Gemini 2.5 Flash Image model: fast, cheap, but text inside images drifts and output tops out around 2K. “Nano Banana Pro” is Gemini 3 Pro Image (GA June 2026): it reasons through complex edits, renders legible in-image text, outputs up to 4K, and holds up to five consistent people. Use Pro for hero assets; use the Flash models for fast iteration.
Q: How do I make it change only one thing and keep the rest? A: Name the keep-list explicitly — “change the shirt to navy; keep the face, pose, background, and lighting unchanged” — and make one change per turn. Vague prompts let the model redraw things you didn’t ask it to touch.
Q: Is Nano Banana free, and is there a watermark? A: The Gemini free tier allows a small number of edits per day at 1K resolution, and free outputs carry a visible Gemini sparkle watermark plus an invisible SynthID signature. Paid plans (Google AI Plus/Pro/Ultra) and API outputs drop the visible watermark and keep only the invisible SynthID. Higher resolutions (2K/4K) and higher daily caps require a paid plan.
Q: How is Nano Banana different from ChatGPT image editing? A: Nano Banana is sharper at targeted local edits that preserve the rest of the image, and Nano Banana Pro is the stronger choice for legible text inside an image. ChatGPT tends to redraw more of the scene on an edit. For “swap this one element only,” Nano Banana usually wins.
Related
- Gemini image generation
- AI image prompt basics
- AI image style drift fix
- AI consistent character images
- AI image aspect ratio guide
Tags: #Tutorial #Image generation