Imagen 3 is Google’s AI picture generator, which was introduced again in Might on the firm’s I/O developer convention. It launched in a restricted capability within the US in August however turned out there to free Gemini customers final month. I’ve been utilizing it ever since to create all kinds of pictures, and whereas it’s a formidable software general, it does have a number of limitations that hinder the general expertise.
Right here’s the place Imagen 3 struggles
The primary restrict to pay attention to is which you can’t generate pictures of individuals, at the very least with a free Gemini account. This doesn’t simply apply to creating pictures of well-known folks, which not many image-generating instruments permit for anyway, however folks typically. So a immediate like, “create a picture of two random folks dancing” won’t return any outcomes. For reference, ChatGPT additionally has this restrict in place for its free tier.
You possibly can create pictures of individuals for those who improve to Gemini Superior.
Nevertheless, you may create pictures of individuals — excluding well-known ones — for those who go for a Gemini Superior subscription. I attempted it out, and it’s a hit-and-miss. Whereas it might probably generate pictures which are so sensible it’s onerous to inform whether or not they’re AI-generated or not, typically the outcomes it produces are subpar. Take a look at the 2 examples under. The one on the left comes throughout as very sensible and appears prefer it was taken by knowledgeable photographer, whereas the opposite one simply appears to be like cartoonish. Even when prompting the software to make the picture extra sensible a number of occasions, the modifications it made have been minimal.
Talking {of professional} photographers, let’s transfer on to the second restrict or difficulty I see with Imagen 3. Even when producing a sensible picture, whether or not of an individual, an animal, or an object, the consequence appears to be like skilled as a substitute of informal. Each picture is picture-perfect, with the bokeh impact steadily added to make it look extra interesting. Each image Think about 3 creates appears to be like prefer it was closely edited, which is ok if that’s the look you’re going for, however being able to make pictures look extra informal can be nice.
I feel the perfect pictures are typically those which are uncooked. The unedited ones you took with out a lot thought when the lighting wasn’t good and the folks you captured didn’t even know you snapped a photograph. That’s the place Think about 3 struggles, though it’s value mentioning that that is true for nearly each AI picture generator on the market.
This brings me to the third main difficulty with Imagen, which is modifying the pictures created. If I create a humorous picture of a cat sporting a hat and consuming a popsicle after which wish to edit it with a further immediate, Imagen 3 will create a model new picture in Gemini. So, for instance, if I just like the picture created however simply wish to change the colour of the hat from black to blue, the software will generate a brand new picture altogether and alter the colour of the hat as a substitute of simply altering the hat’s coloration and leaving every little thing else as is. Granted, the brand new picture does look comparatively just like the previous one when utilizing the correct immediate, but it surely’s nonetheless not the identical, which isn’t superb. This makes it unimaginable to edit an image to perfection, particularly with a number of prompts that can generate a brand new picture each time. Take a look at the instance under and see for your self.
One other difficulty is that I can’t change the side ratio. Photos are created in a 1:1 side ratio by default and might’t be modified. If I immediate the software to vary it to 16:9, Gemini simply says it’s going to however then generates a brand new picture with the identical side ratio. Nevertheless, it appears to be like like this may change quickly, as the flexibility to vary the side ratio is already within the works.
Limits apart, Imagen 3 is nice
Let me simply make it clear that I’m not making an attempt to bash Google’s fancy AI picture generator. I simply wish to spotlight the boundaries I bumped into whereas testing it in order that you realize what to anticipate. Limits apart, Imagen 3 is definitely a really spectacular software. I’ve tried out just a few of its rivals as properly, and whereas every AI picture generator has its execs and cons, I’d say Imagen 3 is among the many finest ones on the market. My colleague Calvin agrees. He in contrast the software in opposition to rivals and located that it’s the perfect one on the market when it comes to high quality.
We’re nonetheless within the early phases of AI-generated content material.
When Imagen 3 simply will get it proper, the outcomes are excellent. Photos of animals, cities, folks, and the rest for that matter come out nice — for those who can stay with a photoshopped look. Don’t take my phrase for it. Check out the gallery under to see for your self. And needless to say we’re nonetheless within the early phases of AI-generated content material, so simply think about what the software program will be capable to do just a few years down the road.
Different limits to pay attention to
These are the boundaries I got here throughout whereas testing the software and didn’t anticipate — apart from the shortcoming to generate pictures of individuals as a free consumer — though there are different limits in place Google clearly states on its web site. It’s value itemizing them out in order that you realize what to anticipate.
Imagen 3 won’t create a picture it deems inappropriate, even with a paid plan. That features photos associated to violence, harassment, intercourse, discrimination, and the likes. This additionally goes for pictures that encourage harmful exercise and people with dangerous factual inaccuracies that will pose a danger to somebody’s security.
These are all acceptable limits, and a lot of the large AI image-generating instruments have them in place, not counting FLUX.1 utilized by Grok.
