Nano Banana 2 Image Editing Guide: How to Modify Existing Photos with Prompts

Nano Banana 2 — Promptyze’s nickname for Google’s Gemini 3.1 Flash Image generator — isn’t just a text-to-image tool anymore. The big upgrade that landed in late February 2026 is image editing: you can feed it an existing photo and reshape it with plain English prompts. No sliders, no masks, no Photoshop subscription required. You describe what you want changed, and the model figures out the rest.

That sounds simple. And sometimes it is. But “change the background to a Japanese garden” and “remove the power line in the top-right corner” behave very differently under the hood, and knowing which prompts actually work — and which ones make the model hallucinate a nightmare — saves a lot of frustration. This tutorial walks through the full editing workflow, from uploading your first image to getting consistent results across a multi-character scene.

What You’ll Achieve

By the end of this guide you’ll know how to use Nano Banana 2 to replace backgrounds, relight subjects, change clothing or object colors, add and remove elements, apply style transfers, and render legible text onto existing images. You’ll also have a set of copy-paste prompts that work reliably, plus the protips that separate clean edits from messy ones.

Requirements

You need access to the Gemini app (gemini.google.com), Google AI Studio (aistudio.google.com), or the Gemini API via Vertex AI or Antigravity. The image editing capability in Nano Banana 2 runs on Gemini Flash Image — the same model behind text-to-image, now extended to accept an image as input alongside your text prompt. A Google account gets you into the Gemini app immediately. For API access, you’ll need a project in AI Studio or Vertex AI with the Gemini 3.1 Flash Image model selected. The model stamps every output with a SynthID watermark, so keep that in mind if you’re publishing commercially.

Step 1 — Upload Your Image

In the Gemini app, click the image icon in the prompt bar and upload your photo. JPEG and PNG work without friction; very large RAW files may need a quick export first. In AI Studio, use the “Add media” button in the prompt panel and select your image. Via the API, you pass the image as a base64-encoded inline part or as a file URI — the Gemini API documentation covers both methods if you’re going that route through Antigravity or direct API calls.

One thing worth knowing early: the model treats your uploaded image as context, not as a locked canvas. It will reinterpret the scene rather than pixel-pushing specific regions. That’s a feature for style transfers and relighting, and a liability when you need surgical precision. Plan your prompts accordingly.

Step 2 — Describe the Edit, Not the Result

The most common beginner mistake is describing the final image instead of the transformation. “A woman in a red dress standing in Paris” when your input shows her in a blue dress in Chicago is ambiguous — the model doesn’t know what to keep. Instead, anchor your prompt to the change: “Change her dress from blue to red. Keep everything else identical.” That second sentence does a lot of work.

Pro tip ✅

Add “preserve all other details” or “keep the subject, lighting, and composition unchanged” to the end of any edit prompt. It’s the closest thing to a lock layer in plain English, and it dramatically reduces unwanted drift in areas you didn’t touch.

Here are the core editing prompts, organized by use case. Each one is copy-paste ready — upload your image, paste the prompt, generate.

Background Replacement

Background swaps are where Nano Banana 2 genuinely earns its keep. The model is good at edge detection around hair and complex shapes, though very fine hair against a busy background still trips it up occasionally.

Replace the background with a foggy Tokyo street at night, neon signs reflecting on wet pavement. Keep the subject's position, lighting direction, and all foreground details exactly as they are.

This works because it gives the model a vivid replacement scene while explicitly protecting the foreground. The more specific your replacement scene, the less the model improvises in ways you don’t want.

Remove the background entirely and replace it with a clean white studio backdrop with soft even lighting. Do not alter the subject in any way.

Product photographers: this one is your friend. Clean white background, no shadows, product stays untouched. Run it, then export for your e-commerce listings.

Relighting and Color Grading

This is where Nano Banana 2 pulls ahead of simpler editors. It doesn’t just tint the image — it recalculates how light falls on the subject based on your description.

Relight the scene with warm golden hour sunlight coming from the left side. Add soft shadow falling to the right. Keep the subject's features and clothing unchanged.

Pair this with a portrait and you can turn a flat midday shot into something that looks like it was taken during the best hour of the day.

Apply a muted, desaturated film look: reduce color saturation by about 40%, add slight grain, lift the shadows slightly. Do not change any subjects or objects in the scene.

The percentage instruction is a useful trick — the model interprets approximate numerical guidance and applies it more conservatively than if you just said “desaturate.”

Pro tip ✅

When relighting, always specify the light direction (left, right, above, behind) and the quality (soft, harsh, diffused, directional). “Add better lighting” is meaningless to the model. “Add soft diffused light from above-left” gives it something to work with.

Object and Element Editing

Removing objects, adding props, changing colors — these are the everyday editing tasks, and they’re now faster than opening Photoshop.

Remove the person standing in the background on the right side of the frame. Fill the area naturally to match the surrounding environment.

The “fill naturally” instruction is important here. Without it, the model sometimes replaces the removed element with something plausible but wrong — a tree where there wasn’t one, for instance.

Change the color of the car in the foreground from silver to matte black. Keep the reflections, shadows, and all other elements in the image unchanged.

Color changes on objects with reflections are tricky because reflections need to update too. Explicitly mentioning reflections in your prompt nudges the model to handle them correctly rather than leaving silver highlights on a supposedly black car.

Add a coffee cup on the desk in front of the laptop. Match the lighting and perspective of the scene. Keep everything else unchanged.

Adding objects works best when you specify placement, lighting match, and perspective. Leave any of those out and you’ll get an object that floats unconvincingly.

Warning ⚠️

Removing people — even in backgrounds — can trigger content policy checks depending on context. If your edit gets flagged, try rephrasing: “Fill the background area on the right to show only the street scene, matching the existing environment” tends to get further than “remove the person.”

Style Transfer

Style transfer is where Nano Banana 2 gets genuinely fun. You’re not filtering the image — you’re asking the model to redraw it in a different artistic register while keeping the underlying composition.

Redraw this photograph in the style of a 1960s vintage travel poster. Use flat colors, bold outlines, and a limited palette of warm oranges, yellows, and blues. Maintain the original composition and subject placement.

The key to style transfer prompts is describing the style through its visual characteristics, not just its name. “Vintage travel poster” alone produces inconsistent results. Specifying flat colors, bold outlines, and a color palette forces a more coherent interpretation.

Convert this photo to a watercolor painting. Use loose, flowing brushstrokes and allow colors to bleed slightly at the edges. Preserve the subject's recognizable features and the overall scene layout.

Pro tip ✅

If you want a style transfer that keeps the subject looking like a specific real person or product, add “maintain the subject’s exact facial features and likeness” or “maintain product shape and branding details exactly” — without this, the model often abstracts features into the new style, which looks great artistically but not if accuracy matters.

Text Rendering on Images

One of the genuine improvements in Nano Banana 2 is text rendering — adding legible, well-placed text to an existing image. Previous versions of Gemini image generation had a reputation for producing garbled letters. The current model handles short text strings reliably.

Add the text "SUMMER SALE" in large bold white sans-serif letters centered at the top of the image. The text should sit clearly above the main subject with sufficient contrast against the background.

Place the text "est. 1987" in small italic serif letters in the bottom-right corner of the image, in a warm cream color. Keep all other elements unchanged.

Keep text strings short — under six words — for the most reliable results. Longer strings still have a higher error rate. If you need a full sentence rendered accurately, generate it in sections or use a dedicated text tool for the typography layer.

Note 💡

All images generated or edited by Nano Banana 2 carry a SynthID watermark embedded in the pixel data — invisible to the eye but detectable by Google’s verification tools. This persists through most common edits and exports. Factor this into any commercial publishing workflow.

Subject Consistency Across Multiple Edits

Nano Banana 2 supports subject consistency across sessions involving up to five characters — a feature that matters if you’re building a visual narrative or a product campaign with recurring figures. The way to use this in an editing workflow is to reference the original image explicitly and describe your character in enough detail that the model anchors to them rather than drifting.

This image shows a woman with short red hair, pale skin, and round wire-frame glasses wearing a navy blazer. Edit only the background — replace it with a modern open-plan office environment. Keep the subject's appearance, expression, and position identical to the original.

That level of character description in the prompt acts as a consistency anchor. If you’re doing a sequence of edits across multiple images featuring the same person, repeat that description block in every prompt. It’s verbose, but it keeps the model from subtly drifting the character’s features across iterations.

Pro tip ✅

Build a “character card” as a text snippet you paste into every prompt involving that subject: name, hair color, skin tone, distinctive features, what they’re wearing. Takes thirty seconds to write once, saves you five rounds of regeneration every time you start a new edit.

The Editing Workflow in AI Studio vs. the Gemini App

The Gemini app is the faster option for one-off edits — upload, prompt, download, done. AI Studio gives you more control: you can adjust the temperature parameter (lower values like 0.4–0.6 produce more conservative, faithful edits; higher values like 0.9–1.0 give you more creative interpretation), and you can inspect the API response directly. For production workflows or batch editing via the API, AI Studio is where you set up and test your prompts before handing them off to a script. Vertex AI adds enterprise-grade access controls and regional data residency if that matters for your use case.

Pro tip ✅

In AI Studio, set temperature to 0.4 when you need precise, faithful edits (background removal, color changes, object removal). Bump it to 0.8–1.0 for style transfers and creative reinterpretations where some improvisation is welcome. The Gemini app doesn’t expose this control, so if precision matters, go to AI Studio.

What Actually Works — and What to Avoid

Nano Banana 2 handles backgrounds, relighting, color changes, and style transfers with consistent results. Text rendering is solid for short strings. Subject consistency is reliable if you anchor your prompts with detailed character descriptions. Where it still struggles: very fine hair against complex backgrounds (some fringing is common), precise geometric changes (“rotate the cup handle 90 degrees” produces unpredictable results), and any edit that requires understanding three-dimensional spatial relationships the model can’t infer from a flat image.

Avoid 🚫

Don’t use Nano Banana 2 for edits that require pixel-level precision — retouching skin texture, adjusting specific facial features, or removing very small objects in complex areas. For that kind of surgical work, a traditional editor or a dedicated inpainting tool is still more reliable. Nano Banana 2 is a scene-level editor, not a pixel-level one.

Start Editing

The fastest way to understand what Nano Banana 2 can do is to run the prompts above against a photo you actually care about. Background replacement and relighting will probably surprise you — they’re genuinely good. Text rendering will impress you if you’ve been burned by earlier AI image generators. And once you build the habit of anchoring prompts with “keep everything else unchanged,” the results get much more predictable. The model is capable; the prompts just need to be specific enough to direct it. Now you have the prompts. Go use them.

promptyze

See Full Bio