Meta’s AI-Powered Image Generator ‘Imagine’ Trained on Facebook & IG

Meta, the company behind Facebook and Instagram, made headlines on Wednesday with the launch of its free standalone AI-powered image generation website known as “Imagine with Meta AI.” Powered by the Emu image-synthesis model, this innovative platform has been trained using a staggering 1.1 billion images from the public domain of both Facebook and Instagram.

A Creative Playground — Imagine with Meta AI

Imagine with Meta AI follows in the footsteps of other AI image generators, such as Stable Diffusion, DALL-E 3, and Midjourney. However, it comes with its unique capabilities. A Meta account is all you need to use the website, and you can create this account using your existing Facebook or Instagram profiles. Once in, your text prompts will be transformed into four unique images, each measuring 1280×1280 pixels, which you can then save in JPEG format. All the images carry a small “Imagined with AI” watermark logo in the lower left-hand corner.

The AI Experiment

Ars Technica put this new AI image generator to the test with low-stakes, informal prompts like “Barbarian with a CRT” and “Cat with a beer” and was pleasantly surprised with the aesthetically unique results. They also conducted adversarial testing. While the generator was designed to filter out violence, explicit language, and references to celebrities and historical figures, it did allow images of commercial characters like Elmo and Mickey Mouse, albeit within certain limits.

The AI has shown proficiency in creating photorealistic images, though not as proficiently as Midjourney. It handles complex prompts better than Stable Diffusion XL, but perhaps not as well as DALL-E 3. Its handling of text rendering could use some improvements, and it gave mixed results when asked to create images in different media outputs like watercolors, embroidery, or pen-and-ink drawings.

The Driving Force — Emu

Emu, the AI model powering Meta’s new features, creates high-quality images via a process called “quality-tuning,” as described in a research paper released by Meta. Emu emphasizes “aesthetic alignment” using a comparatively small but visually appealing set of images, post-pre-training. The large pre-training dataset, consisting of 1.1 billion text-image pairs pulled from Facebook and Instagram, forms the backbone of Emu.

What sets Meta apart from other AI companies in this arena is its vast trove of image and caption data from its platforms. Unlike others who use images scraped from the internet or licensed from commercial stock image libraries, Meta leverages its own cache of user-shared content.

The Ethical Angle

While the utility and application of AI image synthesis models is exciting, there’s a growing concern about the potential misuse of this technology for creating disinformation or harmful content. Meta aims to mitigate such risks by implementing filters and a soon-to-be-activated watermarking system. A small disclaimer also cautions that the images generated might be inaccurate or inappropriate.