Google Gemini has upgraded its AI image editing capabilities in the Gemini mobile app, introducing a handful of entertaining new tools.
The additional features are courtesy of Gemini’s improved consistency when it processes your images, so if you change one aspect of an image, then Gemini is much less likely to mess around with the rest of it, which has been a constant problem with AI images in the past.
To use the new features you don’t need to upgrade the Gemini app, just upload a photo of yourself and try ask Gemini to do what I did below.
Style shift
The first example from Google that caught my eye was that I could reinvent my look without the result being incredibly obviously not me, or at best, a very obvious transfer of my face onto an AI setting.
So I took a headshot and a Google suggestion to ask Gemini to make me a character from a ’90s sitcom. Gemini answered the challenge without changing my face or shirt, but with a big smile, acid-washed jeans, and blindingly white sneakers, not to mention a classic den of a ’90s sitcom for me to stand in.
Photo Blending
Next I tried photo blending, which is broadly similar, but uses two images to make a new scene. I decided to see how it would perform when putting me and a celebrity together.
I uploaded my headshot and an old photo of Elvis, and asked Gemini to have us both play guitars. I’d mark this one down a little for making both faces look like their respective origin photos, but there’s no denying that a casual glance makes it look like Elvis and me are having a good time jamming together.
Multi-turn
Extending Gemini’s ability to hold onto the key bits of a photo is the multi-turn editing option, which lets you modify an image step-by-step.
I used my Elvis duet photo and chose to make our private jam a public concert by asking Gemini to put us outside with a full crowd. I suspect the AI database has some information about the kinds of people who really loved Elvis, judging from the fact that the vast majority appear to be women.
Why the crowd is behind us and not in front of us is less clear.
Style transfer
The last new feature I tested is Gemini’s “mix-up,” which does a style transfer through images. Google suggested rainboots with a flower petal texture, but I went with a beautiful sunset image and asked for a dress with all the colors of that sky. The model on a runway was extra, but it makes sense to me.
I don’t think anyone would be truly fooled by the AI image editing if they pay attention at all. But, of course, better prompts and experimenting might lead to much better examples that might genuinely seem to alter reality.
Hence, the SynthID watermarking system is applied by Google to every AI-edited or AI-generated image in Gemini. There’s both the watermark you can see and an invisible one embedded in the image file.
It won’t stop misuse entirely, but it gives people a chance to spot AI-made or modified visuals. Still, besides the malicious potential, there are some creative opportunities to be had here, just ask my good friend Elvis.
