Nano Banana is the informal name people use for Google's Gemini image-generation and editing models. The useful part is not the meme name. It is that these models can take an existing image and follow editing instructions: change a background, add or remove an object, restyle a scene, or make a new variation from a reference.
That makes Nano Banana interesting inside an image editor, where a generated result is a starting point instead of the final file. In PhotoFresco, AI output becomes a layer you can move, mask, blend, crop and export with the rest of the document.
What Nano Banana is good at
Use it for the broad, ambiguous work that would take a long time to build by hand:
- Creating a new scene around an existing subject.
- Trying several art-direction routes before committing to one.
- Replacing an ordinary background with a believable environment.
- Making a product image feel more like a campaign image.
- Turning a rough visual brief into editable source material.
The best prompt describes both the change and the things that must stay fixed. For a product image, name the product, its color, its logo, its camera angle and the material you do not want changed. For a portrait, call out identity, pose and wardrobe. Those constraints will not make a generative model perfectly deterministic, but they give it a much better target.
The tiers, in plain language
There are two practical choices when you see Nano Banana in an image-generation catalog:
| Tier | Use it when | Trade-off |
|---|---|---|
| Nano Banana Lite | You need to explore, iterate, or make a quick social/image-edit pass. | It is the lower-cost route, so subtle texture and small identity details deserve a closer check. |
| Nano Banana 2 | The output will carry more of the final visual direction or needs a higher-fidelity pass. | More expensive per output; it still needs an editor for the final 10%. |
The names and available routes change quickly, so treat a tier as a workflow choice, not a promise of a specific result. Lite is the sketchbook. The fuller tier is the more deliberate pass. Neither replaces the cleanup stage when an edge, label, or exact proportion matters.
The part an image model does not solve
An image model can make a persuasive whole image and still miss the detail that makes it usable: a product label can drift, hair can merge into a background, text can become almost-right, or a generated shadow can conflict with the composition.
That is where layers and masks earn their place. Keep the generated result on its own layer. Mask the part you want, reveal the original where the model changed too much, and use adjustment layers to match contrast and color. You are no longer asking the model to get every pixel right in one shot.
This is also a cleaner way to prompt. Generate the broad scene first. Then keep only the useful region, rather than repeatedly asking a model to preserve the entire document while changing one small corner.
A reliable browser workflow
- Open the original in PhotoFresco and duplicate the base layer.
- Use Nano Banana for the large visual change, with the original visible as the reference.
- Keep the result on a new layer; do not flatten the document.
- Add a layer mask and paint back any feature that must remain exact.
- Match the result with adjustment layers, then export a PNG, JPEG, or PSD.
The final PSD matters. It preserves the editable decision-making: what came from the original, what came from the model, and what you corrected afterward.
Costs should be visible before you generate
Generative image models have real per-image compute cost. PhotoFresco uses credits for that work rather than hiding it inside a subscription: you can see the model and credit cost before running it, buy credits only when you need them, and use the editor without paying for AI when you do not.
For a look at how Nano Banana Lite behaves beside dedicated background-removal and image-editing models on difficult real images, open the PhotoFresco background-removal Model Lab. It publishes the source images, completed outputs and the receipt-backed cost of each completed run.