This article is based on publicly available information at the time of writing. Google may change model versions, quotas, and commercial terms at any time. For the latest details, check the official Gemini image generation page and the Google AI subscriptions page.
What can Gemini actually do with images now?
For years, "AI image generation" basically meant Midjourney, DALL-E, or Stable Diffusion. Google had Imagen, but it was tucked behind Vertex AI — invisible to most users.
That changed at the end of 2024. Google built image generation directly into Gemini under the name Nano Banana. Open Gemini, ask for a picture, and it draws one. No plugins, no separate app.
By 2025-2026, the lineup had iterated into Nano Banana 2 and Nano Banana Pro. They handle a lot more than the original — and crucially, they fix the long-standing pain points of AI image generation: text rendering (especially non-English text), multi-image composition, and character consistency.
This guide covers what versions exist, how to use them, what they're good at, how many images you can make per day, and whether you can use them commercially.
Nano Banana, Nano Banana Pro, and Imagen — how do they relate?
This is where most people get confused. Let's lay out the architecture.
Nano Banana is Google's nickname for the Gemini image model family. It's not a single model — it's a series, with three versions currently in play:
- Nano Banana (officially Gemini 2.5 Flash Image): the first-generation model, launched in 2025
- Nano Banana 2 (Gemini 3.1 Flash Image Preview): the 2026 mainline model — fast and high-quality
- Nano Banana Pro (Gemini 3 Pro Image Preview): the premium tier — 4K output, real-world knowledge integration, and the strongest character consistency and multilingual text rendering
So where does Imagen fit? Imagen was Google's earlier image generation family (Imagen 3, Imagen 4), accessed mainly through APIs. But Google has announced that all Imagen models will be retired on June 24, 2026, with users migrated to the Nano Banana series.
In other words: going forward, image generation on Google = Nano Banana. There's no other option.
How to start: three paths
The fastest ways to get started:
Gemini web (gemini.google.com). Open the Gemini chat box, click "Tools" at the bottom, and pick "Create image." On the right, you can choose the model — "Fast" runs Nano Banana 2, "Pro Thinking" runs Nano Banana Pro. Then just describe what you want.
Gemini App (mobile). Available on iOS and Android, with the same flow as web. Mobile is especially good for "upload a photo and have it edited," since you can pick from your camera roll directly.
Gemini API (developers). To integrate programmatically, grab an API key from Google AI Studio and call models like gemini-3.1-flash-image-preview or gemini-3-pro-image-preview. Full documentation lives at the Google AI developer docs.
For most users, the web app or mobile app is more than enough. The API is for batch generation or product integrations.
Why is everyone hyping the text rendering?
This is Nano Banana Pro's biggest selling point — and the reason it went viral in non-English-speaking markets.
For years, AI image generators butchered any text in images, especially non-Latin scripts. Signs turned into alien glyphs, posters showed scrambled strokes, comic speech bubbles came out as nonsense. The reason: most models were trained heavily on English, and complex scripts (Chinese, Japanese, Korean, Arabic) have far more intricate character structures.
Nano Banana Pro fixed this. You can now write a prompt like:
Draw a Tokyo ramen shop sign that says "らーめん 山田" in Japanese, red background with white text, neon-lit night market style.
And it actually renders the characters correctly. Same goes for Traditional Chinese, Simplified Chinese, Korean, Arabic, and many other scripts. Posters, comic dialogue, T-shirt prints, product packaging — all of these now work for multilingual content.
If you're producing localized marketing assets, regional product designs, or non-English comics, this is a capability competitors don't match yet.
What other killer features are there?
Beyond text rendering, Nano Banana Pro nails several things AI image gen used to struggle with:
Character consistency (same person, different scenes). You can upload up to 5 reference images of a single character, and the model locks in face shape, hairstyle, body type, and clothing style. Then ask it to place that character into a new scene — beach, office, snow mountain — and the identity stays stable. Not "kind of similar" — actually the same character. For comic creators, illustrators, brand mascot designers, and IP content makers, this is a game-changer.
Multi-image composition (up to 14 reference images). You can upload up to 14 reference images at once (6 at high fidelity), and the model integrates them into a single coherent composition. Think: character from image 1 + scene from image 2 + prop from image 3 + art style from image 4. What used to require three or four tools stitched together now happens in one shot.
Conversational editing (don't regenerate, just edit). If a generated image is 80% right, don't generate a new one. Tell Gemini: "Change the background to dusk." "Make the shirt blue." "Remove the person on the right." It understands what you're pointing at and only modifies that part. Faster than regenerating, and uses fewer credits.
4K resolution (Pro only). Nano Banana Pro outputs at 4K, suitable for print, posters, and large-format use. Nano Banana 2 caps at around 1K — fine for web, not enough for print. If you're producing physical assets, use Pro.
Free vs paid: how many images per day?
This is the question everyone wants answered. Google published clear quotas in 2026 (these numbers may change — below reflects April 2026 public information):
| Plan | Nano Banana 2 (Fast) | Nano Banana Pro (Premium) |
|---|---|---|
| Free | ~100/day | ~3-5/day |
| Google AI Pro | ~1,000/day | ~100/day |
| Google AI Ultra | ~1,000/day | ~1,000/day |
Key takeaways:
- The free tier's Nano Banana 2 quota is genuinely generous — 100 images/day is plenty for individual users
- The bottleneck is Pro model access: free users hit the cap after 3-5 images
- If you need Pro heavily (4K printing, accurate text rendering, commercial assets), upgrading is worth it
- Ultra gives you maxed quotas on both models — for daily heavy production work
If you only generate the occasional meme or background swap, free is fine. If you're a designer, marketer, or content creator producing daily, Pro feels noticeably different.
How to write prompts: six common task patterns
The nice thing about conversational prompts is you don't need to learn a "spell syntax." But a few writing habits make a big difference:
1. Generate from scratch
Draw a Shiba Inu wearing a space helmet, sitting on the lunar surface, Earth in the background, photorealistic style, evening light.
Formula: subject + action/pose + scene + style + lighting. The more specific, the better.
2. Edit an existing photo
Upload a photo, then say:
Change the background to a rainy night street in Tokyo. Keep the person unchanged.
Tip: state what to change first, then what to preserve. Otherwise the model takes liberties.
3. Add text (signs, posters, packaging)
Make a coffee shop opening poster with "Spring Grand Opening" written in brushstroke calligraphy, cream background, latte illustration in the corner.
Tip: explicitly name the language and put the text content in quotes — it dramatically improves accuracy.
4. Swap background or lighting
Upload a photo:
Replace the background with a misty mountain at dawn. Change the lighting to soft golden tones. Keep the subject's pose and clothing exactly as they are.
5. Place a character into a new scene
Upload 3-5 reference images of the same character:
This is my character "Mike." Put him in an 80s arcade-style cyberpunk street scene. Keep his hair, outfit, and body type consistent.
6. Composite multiple images into one
Upload 5-10 reference images (character, scene, props):
Take the character from image 1, place them into the scene from image 2, holding the prop from image 3, with the painting style of image 4.
One last tip: if a result isn't quite right, don't regenerate — just keep talking. "Brighten the light." "Smile a bit more." "Remove the tree on the left." Conversational edits are faster than regenerating, and they cost fewer credits.
Can you use the images commercially? The truth about SynthID
This is where a lot of people stall — "Can I sell what Gemini generates?"
Short answer: yes. Google's terms of service explicitly allow commercial use of Gemini-generated images, as long as you follow the content policies (no violence, deception, illegal content, etc.). You can use them for product packaging, list on microstock, or include in client deliverables.
But two things to know:
SynthID invisible watermark (cannot be removed). Every image generated by Gemini — whether through the web, app, AI Studio, Vertex AI, or API — has Google's SynthID embedded in it. It's invisible to the human eye, but Google's own detector can verify whether an image is AI-generated. The Gemini app rolled out an "upload an image and ask if it's AI-generated" feature in November 2025. Google says SynthID typically survives compression, cropping, recoloring, and format conversion, though they also note it can become undetectable after many alterations. SynthID is an AI content marker, not a copyright stamp.
Visible watermark (depends on your plan). Free tier downloads include a small "Google AI" badge in the bottom-right corner. The lower-tier Google AI Pro plan still retains this visible watermark — you need to upgrade to a higher Pro tier or Ultra to remove it. Images generated through the API or inside Google AI Studio don't carry the visible watermark.
On commercial legality. Google's terms explicitly permit commercial use of Gemini-generated images (subject to content policies). Based on current public information, images containing SynthID can still be used commercially — no major platform or law prohibits it. That said, AI content regulations and platform policies are evolving quickly, so re-check the latest rules before any significant commercial deployment.
Three commercial guardrails:
- Don't generate infringing content (known IP, celebrity likenesses, registered trademarks)
- Don't make misleading content (fake news, fake celebrity endorsements)
- Some microstock platforms require AI-generated disclosure — follow their rules
Gemini vs Midjourney vs ChatGPT: how to choose
Quick decision guide:
Pick Gemini Nano Banana if you produce non-English content, need conversational editing, want to brainstorm + generate in the same interface, already use Gemini or Google services, or care about cost (the free tier is generous).
Pick ChatGPT image generation (GPT-Image) if you already use ChatGPT, need tight integration between copy and image generation, or work mostly in English.
Pick Midjourney if you chase peak visual quality, produce gallery-grade illustrations, are happy to learn a prompt syntax, and value the community style references.
For mainstream users — and especially anyone working with non-English text — Gemini is currently the most practical option. The combination of free quota, multilingual text rendering, and permitted commercial use sets a low bar to entry.
FAQ
Why is the text in my Gemini image still gibberish?
Make sure you're using Nano Banana Pro (Gemini 3 Pro Image), not Nano Banana 2. Accurate non-English text rendering is mainly a Pro capability. The free tier only has 3-5 Pro generations per day — once you hit the cap, you fall back to Nano Banana 2, where text quality drops.
Why can't I find Imagen 4 anymore?
Google has announced all Imagen models will be retired on June 24, 2026, with users migrated to the Nano Banana series. If a tutorial is still teaching Imagen 4, it's outdated — learn Nano Banana instead.
How many images can I make per day?
Free: ~100/day on Nano Banana 2 and ~3-5/day on Nano Banana Pro. Google AI Pro gives ~1,000/day Fast + ~100/day Pro. Numbers may shift as Google adjusts — check the official page for current limits.
Is the visible watermark obtrusive on the free tier?
No. It's a small badge in the bottom-right corner. If you want to skip the visible badge entirely, use the API or generate inside AI Studio — SynthID will still be embedded invisibly, but there's no visible mark.
Can I sell Gemini-generated images as NFTs or list them on microstock?
Yes. Google's terms allow commercial use. Practical caveats: (1) no infringing content (no known IP or celebrity likenesses); (2) some microstock platforms require AI-generated disclosure — follow their rules; (3) SynthID will persist in the file but doesn't block sales.
How many reference images can character consistency use?
Nano Banana Pro supports up to 14 reference images, with up to 6 at high fidelity. For a single character, 5 is usually enough to lock in the key traits (face shape, hair, build, outfit).
Related Reading
- Gemini Free vs Paid: Google AI Plus, Pro & Ultra — What's the Difference? — Decide whether upgrading is worth it
- What Are Gemini Gems? A Complete Beginner's Guide — Understand Gems in 5 minutes
- Browse All Featured Gems →