Meta publicly launches AI image generator trained on your Facebook, Instagram photos

Meta Platforms, the parent company of Facebook, Instagram, Whatsapp and Quest VR headsets and creator of leading open source large language model Llama 2 — is getting into the text-to-image AI generator game.

Actually, to clarify: Meta was already in that game via a text-to-image and text-to-sticker generator that was launched within Facebook and Instagram Messengers earlier this year. However, as of this week, the company has launched a standalone text-to-image AI generator service, “Imagine” outside of its messaging platforms.

Meta’s Imagine now a website you can simply visit and begin generating images from: imagine.meta.com. You’ll still need to log in with your Meta or Facebook/Instagram account (I tried Facebook, and it forced me to create a new “Meta account,” but hey — it still worked).

Early reactions are mixed

Already, AI artists around the web are experimenting with Meta Imagine to produce high-quality imagery quickly and consistently, with some comparing it to other popular AI image generators such as Midjourney, Stable Diffusion, and OpenAI’s DALL-E 3.

VentureBeat’s brief, unscientific tests showed that it only sporadically produced realistic human figures and structures — often our imagery included strange glitches like “melted” body parts and scenery.

Imagine’s interface is extremely minimalistic — a prompt field and four generated images the user can download, with no option to remix, though that is available in its messenger apps.

It also has no way of resizing the imagery beyond a 1:1 aspect ratio square, at least not that we saw. Typing in different aspect ratios into the prompt did not change the aspect ratio of the resulting generated images.

All the images also include a watermark baked into the image in the lower right corner reading “Imagine” with a logo of symbols.

A spokesperson from the company wrote to VentureBeat in an email that “in the coming weeks, we’ll add an invisible watermark to increase transparency and traceability of these AI-generated images.”

Still, Meta appears to be attempting to offer a functional, free competitor to existing AI art generators — many of which require paid subscriptions.

That makes sense given the vast trove of training data, particularly of human figures, that Meta had access too when developing its underlying model.

Built atop Emu, trained on user-generated Facebook and Instagram images

Meta’s Imagine service was built on its own AI model called Emu, which was trained on 1.1 billion Facebook and Instagram user photos, as noted by Ars Technica and disclosed in the Emu research paper published by Meta engineers back in September.

An earlier report by Reuters noted that Meta excluded private messages and images that were not shared publicly on its services.

With other rival AI art generators coming under fire — and even facing class-action lawsuits for alleged copyright violation — for scraping artworks from the public web to train their models on, Meta’s decision to train on its user data seems prudent and wise. But it will not stop criticisms that the people posting said photos never intended them to be used in this way.

When developing Emu, Meta’s researchers also fine-tuned it around quality metrics. As they wrote in their paper: “Our key insight is that to effectively perform quality tuning, a surprisingly small amount  — a couple of thousand — exceptionally high-quality images and associated text is enough to make a significant impact on the aesthetics of the generated images without compromising the generality of the model in terms of visual concepts that can be generated. ”

Interestingly, despite Meta’s vocal support for open source AI, neither Emu nor the Imagine by Meta AI service appear to be open source. We’ve reached out to the company to ask why not and if this is coming, and will update when we hear back.

The Imagine rollout is part of a fleet of updates Meta is making across its apps with new AI-enabled features, including a “reimagine” feature that allows users to “long press” on AI generated images within Facebook Messenger and Instagram messages and create new ones based on text prompts.