Meta released a new AI image-generating website called “Imagine with Meta AI” on Wednesday. This tool allows users to design high-quality images by simply describing them in natural language. Imagine that Meta AI is powered by the Emu image-synthesis model, a previously existing model by Meta. The same model is utilized by some of the leading image-generating tools such as OpenAI’s DALL-E, Stable Diffusion, and Midjourney.
This latest image-generating tool has been trained using 1.1 billion publicly visible images on popular social media platforms Facebook and Instagram, which can help render a novel image based on a written prompt. For now, users residing in the United States can access the platform for free and create up to 4 images every prompt.
Meta AI Image Generator
Meta AI’s newest launch Imagine is capable of generating AI images based on the visual concepts learned by the AI model during its training process. This tool is capable of creating stunning and high-quality AI images using text prompts similar to other leading image-generating platforms such as DALL-E, Midjourney, and Stable Diffusion.
To generate your image on Imagine, users need to have a Meta account, which can be imported from your Facebook or Instagram account. With each prompt, you can create up to 4 1280×1280 pixel images that can be saved into JPEG file format. With every generation you will witness a small “Imagined with AI” watermark logo on the lower left corner of your image.
A blog post was published by Meta which stated, “We have enjoyed hearing from people and how the text-to-image generation feature by Meta AI creates fun and creative content in chats.
Today, we are expanding the access to imagine outside of the chat. “While initially, the messaging experience was designed to provide fun, interactive, and playful interactions, users can now generate free images on the website too.”
1.1 billion Facebook and Instagram Images were used to train Imagine with Meta AI
Meta’s Imagine was built using its own AI model known as Emu, which has been trained using 1.1 billion Facebook and Instagram user images which was uncovered in the research paper of Emu published in September by Meta Engineers.
So, if you have a public Instagram account or Facebook account there is a possibility that your shared pictures might have also contributed to the training process of the AI model Emu. It was reported that in 2016, around 95 million photographs were shared on a daily basis on Instagram. Therefore, the dataset utilized by Meta to train the AI model is a subordinate set of the complete photo library.
According to the research paper shared by Meta in September, this AI model uses a process known as “quality-tuning” to design high-resolution images. Unlike other image-generating platforms that utilize text-to-image models trained using extensive image-text combinations, Emu uses “aesthetic alignment” after pre-training, which uses a set of visually appealing images.
The research paper on Emu was the first paper that didn’t disclaim the potential for models to generate reality-warping disinformation or potentially harmful content. This appears to be more of a general acceptance of AI image synthesis modes which are becoming more and more common among individuals.
Meta on adding watermarks
Meta has pledged to add watermarks to the images created using Imagine in the coming weeks to enhance transparency and traceability. According to Meta, the invisible watermarks will be created along with the AI model and detectable with the corresponding Model. As of now, there is no information on if the detection model will be made public at any point.
Watermarks are resilient to standard image manipulations such as resizing, cropping, changes in color, screenshots, sticker overlays, and more. Meta has stated in a post, “We aim to introduce invisible watermarks for AI-generated images in the future.”
