Nano Banana 2.1: Keep the Subject, Change One Noun

Nano Banana 2.1 is Google’s update to Nano Banana 2 and, on the Gemini API, the new high-efficiency default for image generation and conversational editing. The API id is gemini-nano-banana-2.1. Google’s image-generation docs describe it as Flash-level speed with stronger visual quality, text rendering, multi-turn consistency, and Search grounding across 1K, 2K, and 4K. DeepMind’s Flash image page says 2.1 outperforms earlier Nano Banana models on visual design, mask-based editing, subject consistency, and more natural-looking images. The DeepMind model-card index lists the Nano Banana 2.1 card as updated 6 October 2026. That is a routing change, not a new brief. Keep one subject. Change one noun. Leave prices empty. Write the model you actually used into the file name.

Draft the still in FreeGPTImg. If the surface in front of you is still labeled Nano Banana 2, open Nano Banana 2. Use Nano Banana Pro when the layout is dense. Stay on GPT Image 2 when that is the loop you already trust. Upload a locked hero in image to image, keep materials honest in AI product photography, recrop a paid frame in AI social media ads, and refine a text-only pass in AI text-to-image. Do not stamp a GPT Image 2 file as 2.1.

Name 2.1 before you generate

Do not treat “Nano Banana” as one model. Google’s image-generation guide now splits the family like this.

Name you will see	Official model	Best job	Do not treat it as
Nano Banana 2.1	gemini-nano-banana-2.1	New high-volume drafts and one-noun edits on Gemini	A FreeGPTImg default, or a 4K campaign file you have not inspected
Nano Banana 2	Gemini 3.1 Flash Image (gemini-3.1-flash-image)	Existing 2 loops you already reviewed	The recommended new-project default on Google’s API
Nano Banana 2 Lite	Gemini 3.1 Flash Lite Image	Cheap, fast, single-shot drafts	Multi-reference or multi-turn sequential edits
Nano Banana Pro	Gemini 3 Pro Image (gemini-3-pro-image)	Dense layouts, brand-tight stills, localization, precision control	A Flash workhorse, or Adobe’s own Firefly model
Original Nano Banana	Gemini 2.5 Flash Image	Legacy files only	A current menu you should rebuild a catalog on
GPT Image 2	OpenAI GPT Image 2	Briefs that already live in a GPT Image 2 workspace	“Generated with Nano Banana 2.1”

Google now tells new projects to use Nano Banana 2.1 instead of Nano Banana 2. Nano Banana Pro stays the premium choice for the hardest visual jobs. Original Nano Banana remains the legacy 2.5 Flash Image path; Google’s own migration note for that name points toward 2 Lite, not toward relabeling old files as 2.1. All Gemini-generated images include a SynthID watermark. Availability still varies by app, country, and plan.

A thinking pass is not a license to rewrite the product. Gemini 3 image models use a reasoning step for complex prompts. Google says that thinking is on by default in the API and cannot be turned off. The model may generate up to two interim images while it tests composition. The last image in that pass is the one you review. If the interim frames invent a second bottle, you still discard the final file.

ChatGPT Images 2.5 is not this family. Firefly partner-model names are not Adobe models. FreeGPTImg currently exposes Nano Banana 2, Nano Banana Pro, and GPT Image 2 generator pages. Match the label on the screen, not the blog headline.

Pick the model from the job

Write the job in one sentence before you touch a model menu.

1. Keep-subject catalog still. One SKU, even light, empty label panel if type must be exact. Use 2.1 on Gemini when you need many honest variants. Use Pro when the pack is dense. Use GPT Image 2 if that is the generator already open.
2. One-noun reference edit. Keep the bottle, the face, or the mascot. Change only background, crop, or wardrobe. 2.1 is the new Flash workhorse for this loop on Google’s API. Image-to-image is the right tab.
3. Social ad still. Same object, new crop, empty headline band. Prefer speed. Recrop later. Do not stretch a 1:1 catalog shot into 4:5.
4. Text-in-image concept. Short headline, large type, empty price and date fields. Google still says Gemini works best if you generate the words first, then ask for an image that contains those words. Overlay real prices in an editor.
5. Video poster or keyframe. Google documents video-to-image on 2.1, 3.1 Flash Image, and 3.1 Flash Lite Image. That is a Gemini API feature. Do not assume every web generator accepts a YouTube URL.
6. Fast mood board. Many looks, same subject. 2.1 or Nano Banana 2. Lock the subject first. If style eats the product, discard the file.

Search grounding is not a product photo of a stranger. Google’s limitations note says 2.1 and 3.1 Flash Image grounding with Google Search do not support using real-world images of people from web search. Do not prompt “find this athlete and put our hoodie on them.”

Nano Banana 2.1 keep-subject workflow

Use the same sequence in Gemini, a Nano Banana 2 page, Nano Banana Pro, or a GPT Image 2 tab.

1. Name the slot. White catalog still, 4:5 ad, 9:16 story, 16:9 poster, or cutout. One slot.
2. Name the model. gemini-nano-banana-2.1, Nano Banana 2, Nano Banana Pro, or GPT Image 2. Put it in the file name. Relabeling a 2 file as 2.1 is a lie.
3. Lock the hero. Photograph or generate one as-sold still. Even light. Readable silhouette. No extra SKUs.
4. Write keep / change / do-not. Keep silhouette, pack size, color, and label artwork. Change only one noun. Do not add logos you do not own, barcodes, prices, or a second product.
5. Generate three options on the same crop. Changing aspect ratio between option A and option B is a new job.
6. Edit one noun. “Keep this tin. Change only the background to matte seamless white.” If the label warps, stop and revert.
7. Overlay real words last. Prices, dates, URLs, and legal lines never belong in the generator if they must be exact.
8. Inspect at 100%. Check extra fingers, melted logos, fake seals, and whether SynthID or Content Credentials are acceptable for that channel.

Google’s own limits still apply on Gemini: 2.1 and 3.1 Flash Image support character resemblance of up to four characters and fidelity of up to ten objects in one workflow. Ask for a party table of twenty SKUs and the model will invent. Ask for one tin.

Prompt formula

Keep-subject still, not a collage and not a store screenshot.
Model: [Nano Banana 2.1 / Nano Banana 2 / Nano Banana Pro / GPT Image 2].
Subject: [one SKU or one original character], keep [exact color, material, silhouette, label].
Job: [catalog still / one-noun edit / 4:5 ad / 9:16 story / poster concept].
Crop: [1:1 / 4:5 / 9:16 / 16:9], camera [eye-level / 3/4], lighting [soft studio / window].
Change only: [background / crop / wardrobe / empty-band layout].
Text in image: none, or the exact short line “[WORDS]” already drafted, large and correctly spelled.
Do not add: extra SKUs, prices, dates, barcodes, QR codes, seals, celebrity faces, or logos I do not own.
Leave empty space for [headline / price / CTA] that I will add in an editor.

Prompt examples

One-noun white field (2.1 on Gemini, or Nano Banana 2 / GPT Image 2 on FreeGPTImg)

Keep this exact 12 oz matte black soy-candle tin. Change only the background to matte seamless white, RGB 255 255 255.
Do not restyle the product. Do not warp the label. No second tin. No text.

4:5 ad still, empty headline

Create a 4:5 social still of the same black tin on a bare oak table, even window light.
Keep the tin identical. Leave the lower third empty for a headline I will overlay.
No prices, no “October,” no extra candles.

Text first, then the poster

The only words in the image are FALL CANDLE, large, centered, correctly spelled.
9:16 poster of the same 12 oz black tin, empty price band, no date, no percent.
Do not invent a brand name.

Image-to-image, crop only

Image-to-image: keep this exact hoodie, size M, color, and lighting.
Recrop to 9:16 with the garment in the center third. Add only an empty lower band.
No orange dye, no extra garments, no platform UI.

Review the pixels before you spend

Zoom to 100%. A phone grid hides a warped net weight.

• Subject honesty. Same silhouette, pack size, color, and SKU. If a stranger would think this still sells a brand it does not, discard it.
• Model honesty. The file name matches the generator. A Nano Banana 2 page is not gemini-nano-banana-2.1. GPT Image 2 is not 2.1.
• Empty type. Prices, percents, dates, CTAs, barcodes, and retailer wordmarks stay empty of anything you need people to read. Google still prefers “write the words, then ask for the image.”
• People and rights. No recognizable generated faces you cannot clear. No children. No copyrighted characters. Do not use Search grounding to pull a real person’s photo.
• Provenance. Gemini outputs carry SynthID. Downstream apps may keep, strip, or flag credentials. A prettier still is not a license.

Limitations you should not skip

FreeGPTImg continues to offer GPT Image 2 generations with 20 daily free credits and no credit card required. It also exposes Nano Banana Pro and Nano Banana 2 generator pages. That is not a claim that every Google API feature — 4K, Search grounding, mask-based editing, thinking traces, or video-to-image — is the on-site default. Match the model name to the tool you actually used.

Do not assume Adobe’s own-model commercial terms cover a Google partner-model file. Do not promise Amazon, Shopify, or Meta compliance from a prompt. Do not confuse Gemini 3.8 Flash, a language model, with this image family. The image names that matter here are Nano Banana 2.1, Nano Banana 2, Nano Banana 2 Lite, Nano Banana Pro, and original Gemini 2.5 Flash Image.

Google’s limitations still warn that the model will not always return the exact number of images you asked for, that audio is not an image input, and that video input is only documented for 2.1, 3.1 Flash Image, and 3.1 Flash Lite Image. If turn three adds a SKU you do not sell, revert.

FAQ

What is Nano Banana 2.1?

It is Google’s update to Nano Banana 2. On the Gemini API the id is gemini-nano-banana-2.1. Google positions it as the primary high-efficiency workhorse for generation and conversational editing, with Flash-level speed and improved quality, text, multi-turn consistency, and Search grounding across 1K, 2K, and 4K. DeepMind’s Flash image page highlights better visual design, mask-based editing, subject consistency, and more natural-looking images. The model card was updated 6 October 2026.

Should I switch every Nano Banana 2 file to 2.1 today?

No. Keep files that already passed review. Use 2.1 for new Gemini jobs. Regen when a file fails a real job: warped label, wrong crop, or an edit that will not hold. A new default is not a reason to throw away a working catalog shot.

Is Nano Banana 2.1 the same as Nano Banana Pro or GPT Image 2?

No. 2.1 is the Flash-class workhorse. Pro (gemini-3-pro-image) is the premium model for complex, brand-tight work. GPT Image 2 is OpenAI’s model. Use the generator in front of you. Do not stamp a GPT Image 2 file as 2.1.

Can I put prices or a celebrity in the prompt?

Do not put live prices, percents, or dates in the generator if they must be exact. Leave those bands empty and overlay later. Do not prompt copyrighted mascots, celebrity faces, or real people pulled from web search. Google says 2.1 Search grounding does not support real-world images of people from web search.

Does FreeGPTImg run gemini-nano-banana-2.1?

Use the label on the page you opened. FreeGPTImg is a GPT Image 2 workspace with 20 daily free credits and no credit card required, plus live Nano Banana 2 and Nano Banana Pro generator pages. Write that name into the file. Do not invent a 2.1 badge because this article used the Google API id.

Why did the model invent a second product?

The prompt asked for a “set,” a “table,” or “more lifestyle” without locking one SKU. Shorten it. Named pack only. One noun. Generate again. Google documents a four-character / ten-object fidelity band on 2.1; a crowded scene is how you leave that band.

Sources

• Google AI for Developers: Nano Banana image generation
• Google DeepMind: Nano Banana 2.1
• Google DeepMind: Nano Banana 2.1 model card
• Google DeepMind: Model cards
• Google DeepMind: SynthID
• OpenAI: Image generation
• OpenAI: Introducing ChatGPT Images 2.5
• C2PA: Content Credentials

Make one keep-subject still on FreeGPTImg

Write five lines: the SKU, the slot, the one noun you will change, the empty type band, and the model name on the screen. Open FreeGPTImg, draft on Nano Banana 2 or Nano Banana Pro if that is the page in front of you, keep a parallel loop on GPT Image 2, lock the hero in image to image, keep materials honest in AI product photography, recrop ads in AI social media ads, and refine a text-only pass in AI text-to-image. Overlay the real offer only after the still matches the pack you ship.