Meta's new AI image creation tool was trained using 1.1 billion publicly visible photos from Instagram and Facebook.
“Imagine with Meta AI” converts prompts into images, trained using public Facebook data.
See also: Meta: Integrates its AI chatbotinto WhatsApp
On Wednesday, Meta released a free, standalone AI image creation tool, “Imagine with Meta AI,” based on Emu image composition model . Previously, Meta’s version of this technology — using the same data — was only available in messaging and social media apps like Instagram.
If you use Facebook or Instagram, chances are a photo of you (or someone you took) helped train Emu. In a way, the old saying “If you’re not paying for it, you are the product” has taken on a whole new meaning. Still, in 2016, Instagram users were uploading over 95 million photos per day, so the dataset Meta used to train its AI model was a small subset of its overall photo library.
As Meta states that it only uses publicly available photos for training, setting your photos to private on Instagram or Facebook should prevent them from being included in future training of the company's AI model (unless it changes that policy, of course).
See also: Windows 11 KB5032288: Improves Copilot AI and fixes bugs
Similar to Stable Diffusion, DALL-E 3 , and Midjourney, Imagine with Meta AI creates new images based on what the AI model “knows” about visual concepts learned from training data. Creating images using the new site requires a Meta account, which can be imported from an existing Facebook or Instagram account. Each creation produces four 1280×1280 pixel images that can be saved in JPEG format. The images include a small “Imagined with AI” logo in the lower left corner.
“We’ve enjoyed hearing from people about how they’re using imagine, Meta AI’s text-to-image feature, to create fun and creative content in conversations,” Meta says in its press release. “Today, we’re expanding access to imagine outside of conversations, making it available in the United States starting at imagine.meta.com. This standalone experience for creative enthusiasts lets you create images with technology from Emu, our organization’s image model.”
We tested Meta’s new AI image generator through a series of low-risk, informal tests using the image composition protocol and discovered aesthetically innovative results, as you can see above. (As a side note, when creating images of people with Emu, we noticed that many of them looked like typical fashion Instagram posts.)
We also tried our luck in aggressive tests. The generator seems to filter out most scenes of violence, foul language, sexual content and themes, and the names of celebrities and historical figures, but it does allow commercial characters like Elmo and Mickey Mouse.
Meta's model generally produces photorealistic images well, but not as well as Midjourney. It can handle more complex cues better than Stable Diffusion XL, but perhaps not as well as DALL-E 3. It doesn't seem to handle text rendering well in general, and handles various media forms such as watercolors, embroidery, and ink with mixed results. Images of people seem to include a variety of ethnic backgrounds. Overall, it seems about average today when it comes to image composition using artificial intelligence.
See also: Google postpones the launch of Gemini AI to 2024!
As for Emu, the AI model behind Meta’s new image generation features, it’s based on a research paper that Meta released in September. Emu gets its ability to generate high-quality images through a process called “quality tuning.” Unlike traditional text-to-image models trained on a large number of text-to-image pairs, Emu focuses on “aesthetic alignment” after pretraining, using a set of relatively small, but visually appealing images.
At the heart of Emu, however, is the aforementioned massive pre-training dataset of 1.1 billion text-image pairs sourced from Facebook and Instagram. In Emu’s research paper, Meta doesn’t specify exactly where the training data came from, but reports from the Meta Connect 2023 conference say that Meta’s president of global affairs, Nick Clegg, confirmed that they were using social media posts as training data for their AI models, including the images that feed Emu.
This is a change in approach compared to other AI companies, as Meta has access to so much image and caption data from its services. Other image composition models use images illegally obtained from the Internet, licensed from commercial image libraries, or a combined approach.
It is also interesting that Meta’s research paper on Emu is the first to express no doubt about the model’s potential to generate misleading apocalyptic information or potentially harmful content. This seems to reflect a general acceptance (or resignation) of the reality of AI image synthesis models, which are now becoming much more common. Whether this is a good thing or not is an open question.
Still, Meta appears to be addressing the issues of potential harmful effects with filters, a proposed watermarking system that isn’t yet functional (“In the coming weeks, we’ll be adding invisible watermarking to the Imagine with Meta AI experience for increased transparency and traceability,” the company says), and a small disclaimer at the bottom of the site: “Images are and may be inaccurate or inappropriate.”
The images may not be accurate, and they may not even be ethical in the eyes of the anonymous editors of the 1.1 billion images used to train the model. But dare we say it: Creating them can be fun. Of course, depending on your mood and how you view the pace of AI image synthesis, that fun can be balanced by an equal level of anxiety.
Source: arstechnica.com
