Friday, December 8, 2023

You Can Use Meta's New AI Image Generator for Free

If you're interested in trying your hand at AI art—ethics aside—you have no shortage of tools to choose from. From OpenAI's DALL-E, to Bing Image Creator, to Stable Diffusion, you have access to AI art whether you're willing to pay or not. So it might not be all that exciting to hear about another tech company jumping into the AI image generator game, but, hey, this train isn't slowing down anytime soon: Meta's just the latest to do it.

How to use Meta's new AI image generator

Meta launched a free AI image generator on Wednesday, following initial AI image generation features in Messenger, Instagram, and WhatsApp. When you fire up Meta's AI image generator site, you're greeted by a series of AI images the company made itself: a dog sitting in the ocean; an astronaut cat (a catstronaut, if you will); a tiger wearing a stylish suit. The site has a text field that invites you to describe an image for Meta AI to generate, but if you want to see that creation through, you'll need to log into your Meta account. If you don't have one (a Facebook account alone isn't enough), Meta is more than happy to help you make one. Whether or not you see this service as worth dealing with Meta's privacy practices is entirely up to you.

Once you're in, it's the same deal as any other image generator: Enter your prompt, hit Generate, and let the AI get to to work. It's pretty fast: Meta AI will spit out five different options in a matter of seconds. Perhaps it's even too quick: In my testing, the images the bot produced weren't necessarily impressive. Most AI images have a sort of "uncanny valley" factor to them, but many of Meta's creations seem like bad photoshops.

I first tried asking Meta AI to generate an image of "drinking a Cherry Coke at a movie theater under the sea." While the first result did an okay job following the prompt, the second result really cracked me up: It seems obvious the bot pulled an image of someone drinking a soda from the side of the lid, and crudely combined it with an image of a soda with the straw in a totally different location.

Meta's AI results for the prompt "Drinking a Cherry Coke at a movie theater under the sea"
The straw is right there, ma'am. Credit: Jake Peterson

It also has a really hard time with keyboards. All the laptops I ask it to generate look like they've been melted in a fire, at least when looking at the keys. Check out this image of a "stressed out Christmas tree working on a MacBook Pro in a coffee shop." To Meta's credit, I appreciate the poorly decorated, sprawling wires to represent the tree's stress, but look at that keyboard! No wonder that tree can't get any work done.

Meta's AI results for the prompt "A stressed out Christmas tree working on a MacBook Pro in a coffee shop."
Credit: Jake Peterson

The "bad photoshop" quality of these images might have something to do with its training set: Meta used public images from its Facebook and Instagram users to train this AI, upwards of 1.1 billion images by some reports. If you have public images on either of these accounts, chances are the AI learned how to make "art" from them.

At the very least, Meta slaps a watermark on these images, in an attempt to make clear they were generated using AI. Alas, this watermark is plainly visible, and easily cropped or edited out. It won't be until companies start rolling out invisible watermarks that AI images will be more easily recognizable in the wild. Although, many are easily recognizable, as they're garbage.



from LifeHacker https://ift.tt/QwcIR02

0 comments:

Post a Comment