Meta introduces 'human-like' AI model for creating images
Meta has introduced a "human-like" AI image creation model that can analyze and finish incomplete images more accurately compared to its rivals. Called Image Joint Embedding Predictive Architecture (I-JEPA), it creates an internal model of the outside world to compare "abstract representations of images." This is how even humans learn new concepts. Its technology helps avoid errors common to AI-generated pictures.
What's the idea behind I-JEPA?
Akin to humans, I-JEPA learns background information related to the world while passively observing it. This data is encoded into "digital representations" for access later. The model predicts the "representation of part of an input" like texts/images by using representations of other parts belonging to the same input. This differentiates it from other generative AI models, that "learn" by distorting/removing parts from the input.
How is Meta's AI model superior to others?
Meta claims rival generative AI models try to fill in every bit of missing information, and make mistakes because they lay too much emphasis on irrelevant details. I-JEPA does not suffer from the same problem as it predicts missing information using "abstract prediction targets" which eliminate unnecessary "pixel-level details." It can predict "higher-level" information about unseen regions in pictures, instead of just pixel-level details.
I-JEPA has fared well in multiple tests
As per Meta, I-JEPA has performed exceptionally well on several computer vision benchmarks and proved to be "more computationally efficient" than rival models. The I-JEPA works quite well on its own and does not require a lot of fine-tuning.
Meta is open-sourcing its models
Meta is open-sourcing I-JEPA's model checkpoints and training code, and next, the firm will see whether the model can be successfully used for videos and image-text paired data. The company believes that sharing its models with others will aid the firm in "spurring innovation, spotting safety gaps, and lowering costs." Meta is aggressively pushing to catch up with its rivals in the AI race.
Meta is also readying AI-based consumer products
Separately, Meta is also pushing generative AI facilities into its consumer products, based on text prompts. These include an Instagram tool that can modify user photos and advertisement tools that can create image backgrounds.
