Meta’s new AI image generator underwent training using a dataset of 1.1B Instagram and Facebook pics
Yet another case of ‘you are the product’?
Meta has introduced an AI image generator named "Imagine with Meta AI," derived from its Emu image-synthesis model, trained on a massive dataset of 1.1 billion publicly available Instagram and Facebook images.
Here are a few key points:
This tool creates images based on text prompts and is now accessible via a standalone website, requiring a Meta account. Users can generate four JPEG images per session.
The AI prioritizes aesthetic appeal but, as noted by Benj Edwards in Ars Technica, it displays average performance in comparison to other AI image synthesis models. Notably, Meta's Emu model diverges from conventional approaches by focusing on aesthetic alignment after pre-training with visually attractive but relatively small image sets.
The extensive training dataset raises questions about data ethics prompting Meta to implement filters and a forthcoming watermarking system for transparency.
So what do you think? Was the AI training opportunistic? Abusive? Or perhaps clever?