AI AI Tools Guidebook
AI Tool Tutorials

Nano Banana / Gemini Image Editing Tutorial

Edit existing images via prompts — change specific elements, keep the rest.

Published: Author: AI Productivity Guide Team 🌐 查看中文版本

What this covers

Edit existing images via prompts — change specific elements, keep the rest.

Key tools and concepts:

  • Gemini: Google’s multimodal AI assistant and the underlying model family, deeply integrated with Google Workspace, Search, and the Gemini API / Vertex AI.
  • Nano Banana: the community nickname for Google’s Gemini 2.5 Flash Image image-editing model. The name first surfaced on lmarena.ai before Google confirmed it; today it is natively part of the Gemini family, not a separate product or app.

What Nano Banana is and how it relates to Gemini

Nano Banana

Nano Banana is the informal name for Gemini 2.5 Flash Image, Google’s image-editing model. It started as an anonymous entry on lmarena.ai that beat other editors in blind tests; Google later confirmed it was theirs and folded it into the Gemini family. There is no separate “Nano Banana” UI — when you upload an image to a Gemini chat and ask to change it, this is the model doing the work.

What it does best:

  • Targeted local edits while preserving the rest of the image (swap the jacket color, leave the face untouched).
  • Multi-turn iterative edits in the same chat: “now make the sky orange,” “now zoom in on the face” — the conversation remembers prior edits.
  • Style transfer with subject preservation: turn the same portrait into watercolor / 3D render / line art without losing identity.
  • Identity-preserving variations: the same person across multiple scenes, outfits, or angles.

Sample prompt:

Here is a photo of my dog on a couch.
Replace the couch with a wooden park bench in autumn,
keep the dog, lighting, and pose exactly the same.
Then in a second message: now add soft golden-hour light.

What it cannot do well: detailed text inside images drifts (logos, signs, paragraphs of copy will warp); complex multi-subject scenes lose secondary detail; aggressive style changes can break composition; there are no millimeter-precise inpainting brushes or layer masks.

Gemini

Gemini is Google’s multimodal AI assistant and the underlying model family — it handles text, code, images (generation and editing), and audio in one chat. The same Gemini chat that answers questions and writes code is also where you reach the image editor: upload an image, describe the change, get a result, iterate.

What it does best:

  • One chat surface for text + image-gen + image-edit, no tool-switching.
  • Free, AI Pro, and AI Ultra tiers in the Gemini app, plus a free developer quota in Google AI Studio and paid programmatic access via the Gemini API and Vertex AI.
  • Long context across a session, so a 10-step edit chain stays coherent.
  • Tight integration with Google Workspace, Search, and Drive when you need source images from your own files.

Sample prompt:

Open gemini.google.com, attach product-shot.jpg,
then: "remove the price sticker on the bottle,
keep the label and reflections intact,
and give me a square crop for Instagram."

What it cannot do well: image editing inherits all the Nano Banana limits above; the free tier has lower rate limits and smaller output sizes; some regions still have feature gaps; for true layer / mask precision you still need Photoshop.

Who this is for

Anyone with a base image they need to tweak — product shots, portraits, social posts, mock-ups — who wants natural-language edits instead of opening a desktop editor.

When to reach for it

You need a small change, not a full re-generation.

Step by step

  1. Open gemini.google.com and start a new chat (sign in with the Google account that has the right plan).
  2. Upload the base image via the attach button.
  3. Describe just the change in natural language: “Replace background with a blue gradient, keep subject and lighting.” Be explicit about what to keep.
  4. Iterate in the same chat — one change per turn (“now make the sky orange,” “now zoom the face”) so the model can build on the previous edit.
  5. Save each intermediate version you might want to go back to; the chat keeps a history but downloads do not always preserve full resolution.

Base → one targeted edit → save → next.

Nano Banana vs other image editors

Photoshop generative fill

Photoshop’s generative fill lives inside a full layer-and-mask editor, so it wins when you need millimeter-precise selection, non-destructive layers, or exact color values. Nano Banana is faster and friendlier for whole-image natural-language edits — “make this look like dusk, keep the model’s face” — without ever opening a masking tool. Use Photoshop when the edit is geometric or pixel-precise; use Nano Banana when it is descriptive.

Flux Kontext (Black Forest Labs)

Flux Kontext is another instruction-driven editor and is often stronger on hard non-destructive composite edits — inserting an object that has to match shadow and perspective, for example — but it is slower and lives behind its own API or third-party UIs. Nano Banana is integrated directly in the Gemini chat with conversational multi-turn iteration, which makes it the default for quick work; reach for Flux Kontext when one specific composite edit is failing in Nano Banana.

Seedream-edit (ByteDance)

Seedream-edit is ByteDance’s image editor and is notably strong on Chinese-language prompts and identity preservation for Asian faces — it tends to handle local idiomatic descriptions better than universal models. Nano Banana is more universal and more deeply integrated with everything else you do in Gemini; if your edits are heavily Chinese-prompted and portrait-focused, it is worth keeping Seedream-edit as a second option.

Common mistakes

  • Asking for many changes at once — the model averages them out and you lose control. Iterate one change per turn.
  • Not specifying what to keep — say “keep the subject, lighting, and composition” explicitly, otherwise the model is free to change them.
  • Forgetting the chat is multi-turn — your fifth message builds on edits 1-4, so going back to the original means starting a fresh chat with the original image.
  • Trying to render long text inside the image — logos, signs, paragraphs will warp. Add text in a real editor afterwards.
  • Using Nano Banana for millimeter-precise layout — for exact crop, alignment, or color values, finish in Photoshop.
  • Not saving intermediate iterations — if iteration 7 is worse than iteration 4, you want iteration 4 still on disk.

FAQ

Q: Is Nano Banana a separate app from Gemini? A: No — it’s the community nickname for Gemini 2.5 Flash Image, Google’s image-editing model. When you upload an image to a Gemini chat and ask to edit it, this is the model doing the work. There’s no separate UI.

Q: How precise can Nano Banana edits be? A: It excels at element-level prompts (“change the shirt color to blue, keep everything else”). It’s not pixel-precise — for exact crops, color values, or alignment, finish in Photoshop or Pixelmator.

Q: Why do my iterations sometimes get worse, not better? A: Compound drift. Each edit on top of an edit erodes the original signal. Save intermediates, and if iteration 7 is worse than iteration 4, restart from iteration 4 with a clearer single-step prompt.

Q: How is Nano Banana different from ChatGPT image editing? A: Nano Banana is sharper at targeted local edits with the rest of the image preserved. ChatGPT’s image edits tend to redraw more of the scene. For “swap this element only” tasks, Nano Banana wins; for full re-generations from a reference, either works.

Tags: #Tutorial #Image generation

Related Articles