Maintain your look as you edit
We launched native image editing in the Gemini app earlier this year, and we’ve been working hard to improve it, with particular focus on maintaining a character's likeness from one image to the next. We know that when editing pictures of yourself or people you know well, subtle flaws matter — a depiction that’s "close but not quite the same" doesn’t feel right. That's why our latest update is designed to make photos of your friends, family and even your pets look consistently like themselves, whether you're trying out a 60’s beehive haircut or putting a tutu on your chihuahua.
Just give Gemini a photo to work with, and tell it what you'd like to change to add your unique touch. Gemini lets you combine photos to put yourself in a picture with your pet, change the background of a room to preview new wallpaper or place yourself anywhere in the world you can imagine — all while keeping you, you. Once you're done, you can even upload your edited image back into Gemini to turn your new photo into a fun video.
Bring your vision to life with advanced editing
Here are a few things to try as you explore this new image editing capability:
Give yourself a costume or location change: Upload a photo of a person or pet, and our model will keep their look the same in every image as you place them in new scenarios. Try putting yourself in different outfits or professions, or even see how you’d appear in another decade — all while still looking like you.
Blend photos together: You can now upload multiple photos and blend them together for a brand-new scene. For example, take your photo and another of your dog to create a perfect portrait of you both on the basketball court.
Multi-turn editing: You can keep editing the images Gemini makes — take an empty room, paint the walls, then add a bookshelf, some furniture or a coffee table. Gemini's working with you all along to alter specific parts of an image while preserving the rest.
Design mixing: Apply the style of one image to an object in another. You can take the color and texture of flower petals and apply it to a pair of rainboots, or design a dress using the pattern from a butterfly's wings.
Originally posted https://blog.google/intl/en-mena/product-updates/explore-get-answers/nano-banana-image-editing-in-gemini-just-got-a-major-upgrade/
Description
Comments (30)
This is the best model I've used so far. Between raw generation and image modification, it's extremely powerful. Last year this is the exact kind of model I was expecting when thinking "Imagine how this is going to be in a year from now!" It's only a matter of time before something like this is open source and can be run on consumer grade hardware.
yeahhhh but until its available locally, straight junk
By far the most terrifyingly impressive model I've ever seen. Perfect text, amazingly controlled layout, consistent characters taken from multiple reference images... As I said, terrifyingly impressive
Nano Banana, in my opinion, is the best AI for image generation. I never thought such a POWERFUL tool could appear so quickly. My respect.
Glimpse into a bright future
This model changed all
Is there a way to use this in comfy ui?
Perhaps if you use some sort of node that uses Gemini's API.
Yes, Comfy is full of API nodes now. Use the "Templates" section on the left hand menu and search for Nano Banana in the "Image" Templates category. You'll have to sign in to Comfy to set up API keys for the various services.
This model is the best image editing tool. By using it, image retouching and correction in Photoshop have become almost unnecessary, significantly improving workflow efficiency and expanding creative possibilities. It is truly an outstanding model.
It's the best image editing tool except when It's not lol
nice
minimum vram requirement or minimum model tarjet video
Try googling
512kb minimum
good
Does it allow the following?
1. NSFW content
2. LoRAs and embeddings as SD
No, no NSFW because it's Google, and LoRAs/embeddings are largely redundant - just pass in a couple images of a character and you've got a consistent character, no training required.
If you want comparable results, with some lewd/NSFW ability, check out Seedream 4.5
You can just use a model similar to your style to remove their clothes with inpainting and then gg ez
I mean its Banana Pro, the best image gen out there I think
I have mixed feelings about Nano Banana Pro, will try to explain. It is definitely most powerful AI tool, I ever used. But could it be called model? It is not like trained models we used to, containing some amount of data up to some date. It could render very specific things, no other model I used before was able to. Sure it have some huge trained core, but I guess, for specific stuff Google just use their own search capabilities, and doing some kind of img2img, cause most of those very specific things are very similar to those images I could find in Google search. So compared to local models it is feels like kind of cheating, just using real images. I could be totally wrong, that is just impression I got after using it for a while.
If you look it up you will find that nano banana pro isn't just a simple diffusion model but also uses "search grounding" to overcome its training data limitations, comparable to gpt-image 1.5. Otherwise it would be simply impossible to have a dataset with that many details, as of now at least. So yes, kind of unfair because the architecture is different but on the other hand it is simply impressive what open weight diffusion models know and can output in a very efficient way in 2026 without the need of a datacenter to run it. This tech might be integrated into newer open weight models sometime in the future.
Agreed, this thing is crazy powerful and easy to use. It does feel like cheating.
after seeing what Nanobanana can do can't wait for Macro Potato
Now that there are 3 different versions of the "Nano Banana" model, do these really have to be grouped under the same model id?
These are completely different models and it would make sense to split them up so they can be discussed separately.
It also makes it easier to tag images by simply selecting e.g. the "Nano Banana Pro" model without having to go into the dropdown menu on the "Nano Banana" model to select the correct version.
Oh? So if Nanobanana is good then where is nanobanan 2?
Nannobanana 2-👀👀👀
you're going bananas
whats the point of this? do you have the weights or this is linked to google api?
Got the weights running on our 3dfx Voodoo 2 cluster, under the desk at my feet!
(It's API - we host a lot of API models)
@theally yeah another voodoo 2 brother found, i owned one and also the generation after that with 16mb (mb not gb lol) , anyway the point its a slippery slope trying to normalise closed source models - api only. Its like in Germany when the east said the wont build a wall, next day we've got the GDR... ;)