Meta launches 'human-like' AI for generating images

Meta has launched a new "human-like" AI model for generating images in a move that could ultimately see platforms like Instagram and Facebook shift towards more AI-generated visual content. — Photo: Fabian Sommer/dpa

WASHINGTON: Facebook parent company Meta has made a new artificial intelligence model designed for image generation available to the public.

I-JEPA, Meta's new "human-like" model, comes at a time when image generators like Midjourney are making rapid advances in creating photo-like images from scratch, albeit often with clear errors, notably with human hands.

Meta says the new I-JEPA model has advanced capabilities in analysing and effectively completing unfinished images, surpassing the accuracy of current models.

While the technology is initially expected to be used primarily by specialists and enthusiasts in the field of AI image generation, it's likely that this technology could lead to a number of AI-based features appearing in Meta platforms like on Instagram.

Unlike conventional generative AI models that rely primarily on the neighbouring pixels of an image, I-JEPA uses "an enormous amount of background knowledge about the world." The AI offering of Silicon Valley giant Meta, which owns Facebook, Instagram and WhatsApp, has so far largely been limited to an AI technology called LLaMA, which can power online chatbots.

In a surprise move in February, chief executive Mark Zuckerberg decided to release the AI language model under an open-source licence, granting the public free and open access to the underlying technology. The new I-JEPA model is also to be published as open source.

The platform is based on the designs of Meta's leading AI scientist, Yann LeCun, who envisages a "human-like" thinking capability in AI systems. By incorporating this kind of thinking, I-JEPA aims to correct common errors in AI-generated images.

"AI researchers have tried to devise learning algorithms that capture common sense background knowledge about the world and then encode it into a digital representation the algorithm can access later," Meta says.

Specifically, this means that unlike previous models, the system is not trained to evaluate every pixel of an image. Instead, the AI model focuses on central elements.

For example, instead of analysing the photo of a person pixel by pixel, I-JEPA has been optimised to recognise central areas and thereby conclude that a person has been depicted who should also have only five fingers on one hand. – dpa