How to do it, step by step
- Upload the photo, not a screenshot Attach the original file from your library. A screenshot has already lost a layer of quality and often carries interface elements the model will try to interpret.
- Describe the change and the constraint together Gemini responds well to a stated constraint. "Change the wall colour to warm white, keep her face, clothing and the lighting on her exactly as they are" is a far more reliable instruction than the first half alone.
- Add reference images if a subject must stay recognisable For a person or a product that has to look like itself, upload two or three more photos of it in the same message. Multi-reference input is where this model is strongest.
- Ask for the aspect ratio you actually need Name 4:5, 9:16 or 16:9 in the prompt and it will render at that shape rather than making you crop afterwards.
- Refine in the same conversation, one change at a time Each follow-up message re-renders, so stack changes gradually and stop as soon as it is right. Likeness drifts with every extra round.
- Check the face against the first image, then download Compare the result with the original rather than the previous version, which hides slow drift. Look at the jaw, nose and skin texture before you keep it.
What Gemini is actually running
"Gemini" is the assistant. The image work is done by a family of models Google nicknamed Nano Banana, and knowing which one you are talking to explains most of the odd results people report.
The app default is Nano Banana 2, formally Gemini 3.1 Flash Image, released on 26 February 2026 and rolled out through Gemini, Search and Ads. Above it sits Nano Banana Pro, Gemini 3 Pro Image, the high-fidelity option Google still points to for work that has to be factually precise. Below it, for developers, is Nano Banana 2 Lite, and the original Nano Banana, Gemini 2.5 Flash Image, is now the legacy model.
All of them accept an image plus text and return an edited image, which is the mode Google's own API documentation calls text-and-image-to-image: add, remove or modify elements while preserving the style and context of the original. Nano Banana 2 takes up to fourteen reference images in a single request, split as ten objects and four characters, and Google states it holds character resemblance across up to five characters.
Resolutions run from 512 pixels to 4K on everything except Lite, which is 1K only, and the supported aspect ratios cover the useful set: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9 and 21:9.
What it does better than the alternatives
Text inside images. This is the clearest lead. Google specifically built Nano Banana 2 to render accurate, legible text for mockups and cards, and it will also translate text that already exists inside an image. If your edit involves words in the picture, this is the tool.
Keeping a subject recognisable across shots. Feeding it several photos of the same person or product and asking for a new scene holds together better than it has any right to, because multi-reference consistency was the explicit target of the last two releases.
Iteration speed. The Flash line is built for speed, so the loop of look, adjust, look again is short enough to actually iterate rather than wait. This matters more than raw quality for most real work.
Aspect ratio control. You can ask for 4:5 for a feed post or 9:16 for a story and get it natively rather than cropping afterwards, which is one of the few places a generative editor genuinely beats a normal one.
A usable free tier. Image generation and editing work without a paid plan, subject to a daily allowance; regenerating with Nano Banana Pro is what the paid Google AI plans unlock. Google adjusts these allowances often, so treat any specific number you read as a snapshot.
Where it falls down
It is still generating, not editing. Same caveat as every tool in this category. The returned file is a new image. Fine grain, skin texture and the small imperfections that make a photo read as a photograph are re-drawn, and on portraits that is visible if you look at the pores and the flyaway hairs rather than the overall impression.
Faces of real people are the hard case. Requests that read as identity modification of a real, identifiable person are where refusals cluster, and Google tightened this area through 2026 in response to deepfake pressure. Expect inconsistency: the same prompt can pass one day and fail the next.
Likeness drift over several rounds. Five characters held across references is the claim, and it is roughly true for one round. By the fourth or fifth sequential edit of a face, small changes accumulate: the jaw narrows, the nose straightens, the skin evens out. Compare against the first image, not the previous one.
Everything is watermarked. All generated images carry a SynthID watermark. It is invisible, it is not removable in any honest sense, and it is designed to survive editing. If you need an unmarked file, this is not the route.
Resolution is not your camera's resolution. 4K sounds generous until you remember that a recent phone shoots considerably more than that, and that the extra pixels you get back were invented rather than recorded.
The rules: age, quotas and SynthID
Three restrictions catch people out, and none of them are obvious until you hit one.
Image editing is 18 plus. Google's Gemini Apps help pages set image generation at 13 or over, but image editing is restricted to users aged 18 and above. Teenagers can make a picture and cannot edit one, which surprises almost everybody who runs into it.
There is a daily allowance. Free accounts get a set number of images a day, paid Google AI plans get more, and the top tier gets a great deal more. Running out of the day's Nano Banana 2 allowance also stops you regenerating anything with Nano Banana Pro. Google changes these figures regularly, so check the current limits page rather than trusting a blog post.
SynthID is on everything. Google embeds an imperceptible watermark in every generated image so that its own tools, and increasingly other people's, can identify the file as AI-made. This is a feature, not a restriction, but it is worth knowing before you use a generated image somewhere that cares.
Images can also be removed for breaching Google's terms or prohibited use policy, which is the formal way of saying the filter applies after the fact as well as before it.
Gemini app, Google Photos or your phone editor
Google now offers conversational editing in two places, and the built-in editor on your phone is still the right answer for a surprising share of jobs.
| The job | Best place | Why |
|---|---|---|
| Change a background or a scene | Gemini app | Full generative control and multi-reference input |
| Put readable words into an image | Gemini app | Text rendering is the model's standout strength |
| Remove an object from a real photo | Google Photos | Magic Eraser works on your file at its own resolution, and the edit stays in your library |
| Tidy a slightly soft old photo | Google Photos | Photo Unblur is free to all users and less aggressive than a full redraw |
| Exposure, white balance, crop | Your phone's editor | Exact, reversible, instant, no upload, no quota |
| Retouch a face | A dedicated retouch app | Generative skin loses the texture that makes a portrait read as real |
| Make a mockup or a card | Gemini app | Native aspect ratios plus legible text in one step |
The pattern: if the answer requires inventing something that was never in front of the lens, use Gemini. If it requires respecting what was, use an editor.
Where ShutterAI fits
Being plain about our own app: ShutterAI is a camera, not a chat assistant, and it deliberately does the opposite job. Glow retouches skin while keeping freckles and texture, True to You fixes skin-tone white balance, Clean Up removes people, and Vault keeps every original so an edit can be undone months later. None of it invents a scene, and none of it replaces what Gemini is good at.
It has not shipped. ShutterAI arrives on iPhone this autumn, free to start, and the early list is on the homepage.
Questions people ask
How do I edit photos with Gemini?
Open the Gemini app, upload the photo, and type the change you want in plain words. It runs Nano Banana 2 by default, returns a re-rendered image, and lets you refine it in follow-up messages in the same conversation.
Is Gemini photo editing free?
There is a free tier with a daily image allowance. Paid Google AI plans raise that allowance and unlock regenerating an image with the higher-quality Nano Banana Pro model. The exact daily figures change often.
Why will Gemini not edit my photo?
The most common reasons are the 18 plus age requirement on image editing, a daily quota that has run out, or a safety filter triggered by an edit that reads as changing a real person's identity.
Does Gemini watermark the photos it edits?
Yes. Every generated image carries an invisible SynthID watermark from Google DeepMind, designed to identify the file as AI-made and to survive normal editing. There is no supported way to remove it.
Is Gemini better than ChatGPT for editing photos?
For text inside an image, for aspect ratio control and for keeping a subject consistent across several references, Gemini is currently ahead. ChatGPT's localised select tool is better when you want one small region changed.