Meta has launched “Imagine with Meta AI,” a free standalone AI image-generator website powered by its Emu image-synthesis model. This model, trained on 1.1 billion publicly visible Facebook and Instagram images, allows users to generate novel images based on written prompts. The release expands Meta’s technology beyond messaging and social networking apps like Instagram.
The vast dataset used to train Emu raises privacy concerns, as images from Facebook and Instagram users contributed to its development. Users who wish to prevent their photos from being included in future AI model training can set their profiles to private.
Imagine with Meta AI is similar to other AI image generators like Stable Diffusion, DALL-E 3, and Midjourney, relying on visual concepts learned from training data to create new images. Users need a Meta account, which can be imported from Facebook or Instagram, to generate four 1280×1280 pixel images saved in JPEG format. The images include a watermark that reads “Imagined with AI.”
In informal tests, the generator produced aesthetically novel results, with a filtering mechanism that screens out violence, curse words, and certain names but allows commercial characters like Elmo and Mickey Mouse. While Emu creates photorealistic images reasonably well, its performance varies across different media outputs, and it struggles with text rendering.
Emu’s unique approach involves “quality-tuning” rather than relying solely on large image-text pairs. The research paper released by Meta in September indicates that Emu focuses on “aesthetic alignment” using a set of visually appealing but relatively small images. The model’s massive pre-training dataset comprises 1.1 billion text-image pairs sourced from Facebook and Instagram posts.
Unlike other AI companies, Meta has the advantage of using its vast image and caption data from its services, eliminating the need to scrape the internet or license images from commercial stock libraries. Notably, Meta’s research paper lacks disclaimers about the potential for the model to generate harmful or disinformation content, reflecting the increasing acceptance of AI image synthesis models.
To address concerns about harmful outputs, Meta has implemented filters, proposed watermarking, and a disclaimer on the Imagine with Meta AI website. The company plans to add invisible watermarking for increased transparency and traceability in the future. However, questions remain about the ethical implications of AI image synthesis and the balance between fun and potential concerns raised by these advancements.