Sign up for our daily and weekly newsletters to stay updated with the latest on industry-leading AI developments.
Meta Platforms, the company behind Facebook, Instagram, WhatsApp, and Quest VR headsets, and creator of the leading open source large language model Llama 2, is entering the text-to-image AI generator arena.
To be clear, Meta was already involved through a text-to-image and text-to-sticker generator launched on Facebook and Instagram Messengers earlier this year. However, as of this week, Meta has introduced “Imagine,” a standalone text-to-image AI generator outside of its messaging platforms.
Meta’s Imagine is now accessible via imagine.meta.com. You’ll need to log in using your Meta or Facebook/Instagram account. I tried with Facebook, which required me to create a new “Meta account,” but it worked.
Initial reactions are varied. AI artists online are using Meta Imagine to create high-quality images quickly and consistently, with some comparing it to other popular AI image generators like Midjourney, Stable Diffusion, and OpenAI’s DALL-E 3.
In recent tests, VentureBeat found that Imagine only occasionally produced realistic human figures and structures. The images often contained strange anomalies like “melted” body parts and backgrounds.
Imagine’s interface is very basic, featuring a prompt field and four generated images available for download. There is no option to remix images outside its messenger apps. Additionally, there are no resizing options beyond the 1:1 aspect ratio square. Inputting different aspect ratios into the prompt did not change the output.
All images include a watermark in the lower right corner that reads “Imagine” with a logo. A Meta spokesperson mentioned in an email to VentureBeat that they plan to add an invisible watermark soon for better transparency and traceability of the AI-generated images.
Meta seems to be aiming to offer a free AI art generator competitor to the existing ones that often require subscriptions, leveraging the vast amount of training data from Facebook and Instagram images.
Meta’s Imagine service is built on their AI model called Emu, trained on 1.1 billion Facebook and Instagram user photos. Ars Technica and Meta’s research paper disclosed this information. According to a Reuters report, Meta excluded private messages and non-public images from the training data.
Other AI art generators have faced backlash and even lawsuits for allegedly using public web-scraped artworks for training, making Meta’s decision to use its user data seem prudent. Nevertheless, there are criticisms since people posting these photos never intended them for such use.
Meta’s researchers fine-tuned Emu based on quality metrics, noting in their paper that just a few thousand exceptionally high-quality images and texts can significantly enhance the aesthetics of generated images without compromising the model’s general visual concept generation capabilities.
Even though Meta supports open source AI, neither Emu nor the Imagine by Meta AI service is open source. We’ve contacted Meta to inquire about this and will provide updates once we hear back.
The rollout of Imagine is part of a series of updates Meta is implementing across its apps. These updates include new AI-enabled features like a “reimagine” function, allowing users to long-press on AI-generated images in Facebook Messenger and Instagram messages to create new ones based on text prompts.
Stay informed by subscribing to our daily newsletter for the latest news. By subscribing, you agree to VentureBeat’s Terms of Service.