Meta opens the gates to its generative AI tech with launch of new Imagine platform

Amongst all the hullabaloo of Google’s Gemini launch, Meta opened the gates to its free-standing image generator website called Imagine with Meta AI.

The company has been tinkering with this technology for some time now. WhatsApp, for instance, has had a beta in-app image generator since August of this year. Accessing the feature required people to have Meta’s app installed on their smartphones. But now with Imagine, all you need is an email address to create an account on the platform. Once in, you’re free to create whatever you want by entering a simple text prompt. It functions similarly to DALL-E. 

We tried out the website ourselves and discovered the AI will create four 1,280 x 1,280 pixel JPEG images that you can download by clicking the three dots in the upper right corner. The option will appear in the drop-down menu.

Below is a series of images we asked the engine to make. You’ll notice in the bottom left corner is a watermark stating that it was created by an AI.

We were surprised to discover that it’s able to create content featuring famous cartoon characters like Homer Simpson and even Mickey Mouse. You’d think there would be restrictions for certain copyrighted material, but apparently not. As impressive as these images may be, there are noticeable flaws. If you look at the Homer Simpson sample, you can see parts of the picture melting into each other. Plus, the character looks downright bizarre.

Limitations (and the work arounds)

A lot of care was put into the development of Imagine. You see, it’s powered by Meta’s proprietary Emu learning model. According to a company research paper from September, Emu was trained on “1.1 billion images”. At the time, no one really knew the source of all this data. However, Nick Clegg, Meta’s president of global affairs, told Reuters it used public Facebook and Instagram posts to train the model. Altogether, over a billion social media accounts were scrapped.

To rein in all this data, Meta implemented some restrictions. The tech keeps things family friendly as it’ll refuse prompts that are violent or sexual nor can they mention a famous person. 

Despite the tech giant’s best efforts, it’s not perfect by any stretch. It appears there is a way to get around said limitations with indirect wording. For example, when we asked Meta AI to create an image of former President Barack Obama, it refused. But, when we entered “a former US president” as the prompt, the AI generated a man that resembled President Obama. 

There are plans to introduce “invisible watermarking… for increased transparency and traceability”, but it’s still weeks away from being released. A lot of damage can be done in that short period. Misuse is something that Meta is concerned about, however, there are still holes. We reached out asking if it aims to implement more protection. This story will be updated at a later time.

Until then, check out TechRadar’s guide on the best AI art generators for the year.

You might also like

• Facebook Messenger gets its biggest ever update – including a major privacy boost
• WhatsApp beta introduces AI assistant to chats – here’s how it works
• How much more data can Meta collect? Probably a lot, thanks to the Meta Quest 3 and Ray-Ban smart glasses