/industry-wired/media/agency_attachments/2024/12/04/2024-12-04t130344212z-iw-new.png)
/industry-wired/media/agency_attachments/2024/12/04/2024-12-04t130332454z-iw-new.jpg)
/industry-wired/media/post_attachments/wp-content/uploads/2023/12/Imagine-with-Meta-AI-Turns-Prompts-into-Images.jpg)
Meta's Emu AI: Imagine and Create!
Meta has launched a free standalone AI image generator website called "Imagine with Meta AI," powered by its Emu image synthesis model. The model, trained on 1.1 billion publicly available Facebook and Instagram images, can generate unique images from written prompts. Previously, this technology was limited to messaging and social networking apps like Instagram. Users may find that their publicly visible Facebook and Instagram images contributed to training the AI.
The new website allows users to create images with a Meta account and generates four 1280×1280 pixel images per session, each marked with a small "Imagined with AI" watermark. Meta expanded access to this feature beyond chats, making it available in the US at imagine.meta.com.
The AI image generator was tested using various prompts, revealing its ability to create aesthetically novel results. Adversarial testing showed that the generator filters out violence, curse words, sexual topics, and certain names but allows commercial characters. The model generally produces photorealistic images but has mixed results with different media outputs.
The Emu model, at the core of Meta's image generation, relies on "quality-tuning" for high-quality image generation. Unlike other AI companies, Meta uses a massive pre-training dataset of 1.1 billion text-image pairs from Facebook and Instagram, where the training data's origin is not explicitly disclosed. Unlike other AI image synthesis models, Meta's Emu research paper does not explicitly address the potential for creating disinformation or harmful content.
To address concerns, Meta employs filters, a proposed watermarking system, and a disclaimer on the website regarding potential inaccuracies or inappropriateness of generated images. Despite ethical considerations, the image generation process is described as fun, raising questions about the balance between enjoyment and potential concerns in the era of widespread AI image synthesis.