Aug 27, 2025 11:06:00

Google releases free ultra-high-quality image editing AI 'Gemini 2.5 Flash Image', which can be instructed in Japanese and can also convert live-action images to anime characters

Google has integrated Gemini 2.5 Flash Image , which has excellent image editing capabilities, into Gemini. Gemini 2.5 Flash Image excels at editing images while preserving their distinctive features, and in tests measuring AI image editing capabilities, it has shown scores far exceeding those of OpenAI's image generation AI and the Flux series.

Google announces native image editing in Gemini app

https://blog.google/products/gemini/updated-image-editing-model/

Gemini 2.5 Flash Image is an AI that can edit an image while preserving its original characteristics by entering 'text explaining the edits' along with the image. Below is an example in which Google CEO Sundar Pichai edited a photo of his pet dog to demonstrate the power of Gemini 2.5 Flash Image. It shows how diverse images can be created while preserving the dog's coat color and body shape.

Our image editing model is now rolling out in @Geminiapp - and yes, it's 🍌🍌. Top of @lmarena 's image edit leaderboard, it's especially good at like maintainingness across different contexts. Check out a few of my dog Jeffree in honor of International Dog Day - though don't let… pic.twitter.com/8Y45DawZBc

— Sundar Pichai (@sundarpichai) August 26, 2025

It's also possible to combine multiple images, and Google has released an example of a photo of a woman holding a basketball and a dog.

The result of the synthesis is as follows. The woman's hairstyle and the dog's fur are retained.

In the '

Image Edit Arena ,' where AI competes in image editing while hiding its true identity, it achieved a score far exceeding that of 'gpt-image-1' and 'flux-1-kontext-max,' taking first place. At the time of testing, it was codenamed 'nano-banana.'

Gemini 2.5 Flash Image is already available to free users in Japan, so I tried it out. This time, I edited the following photo.

Enter the image and edit details into Gemini. In this example, enter 'Change this person's outfit to a yukata.'

An image of a woman wearing a yukata was output in about 10 seconds.

Here's a side-by-side comparison of the before (left) and after (right) images. The hand position and face angle have been slightly altered, but the body shape and facial features have been maintained. The edited image has the Gemini logo in the bottom right, and a digital watermark '

SynthID ' invisible to the human eye is embedded to indicate that it is an AI-generated image.

He then typed, 'Change it to a photo of me smiling and looking up at fireworks on the roof of a building at night.'

I was able to edit it as instructed. If you look closely, you'll notice that the position of the yukata pattern has changed slightly, but at a glance you won't notice the change.

I typed, 'Cut out the character part and change it to a standing image that looks like an SSR character from a social game. In the style of a beautiful illustration from an anime-style game.'

The edited result looks like this. The person has been cut out while still looking like a real-life photo, making it look like a collage image.

I wanted to change it to an illustration style, so I gave the instructions, 'Change it to an anime-style illustration. Have her hold a sword instead of a smartphone, and pose with the sword in hand with a serious expression. Make the character's name 'GIGAko'.'

The result is below. The character's name is displayed at the top of the image, and while the 'GIGA' part was drawn correctly, the 'child' part was distorted. It's also unfortunate that the feet still look like they were taken from a live-action model. While I haven't yet reached the level where I can 'understand even the most general instructions and edit them perfectly,' I can easily get the image I want by repeating the instructions.