Meta is adding about 20 new AI features to Facebook, Messenger, WhatsApp and Instagram. Most of the features are coming to the Meta AI chatbot, such as generating images via text.
The new functionality for the Meta AI chatbot should further expand its capabilities for web search, via Bing, and image generation. For example, Meta AI can now provide more detailed answers to queries from mobile devices, the number of search queries the Meta AI chatbot can process has been expanded, and summarization of Bing search results has been improved.
The generative AI chatbot’s multimodal processing functionality has also been improved. For example, Meta AI can now be used to search clips hosted in the Facebook Reels service.
Imagine image generator
In addition, the built-in Imagine image generator can now upload existing photos and edit them with natural language commands. Imagine will soon also be offered as a free stand alone service. The images generated with this service do receive a watermark invisible to the human eye for greater transparency and traceability for AI-generated images.
Incidentally, initial reactions to the Imagine images are not yet entirely positive. For example, the chatbot’s quality of images would not come close to Midjourney. Nevertheless, these are expected to be improved in the near future.
Emu model with 1.1 billion training images
The Imagine generative AI application is trained on the “Emu” AI model. This model enables the generation of high-quality images through a quality tuning process. Traditional text-to-image models are mainly trained with large amounts of ‘image-text pairs.’ Meta’s EMU model does this differently. This model focuses on ‘aesthetic alignment’ after pre-training the model, using a relatively small set of visually appealing images.
However, the underlying data set is very large. According to Meta, the Emu model for its Imagine AI image generator runs on a training set of as many as 1.1 billion images combined with a ‘lattent diffusion’ architecture. For AI models with such an architecture, the training dataset on which they are built is enhanced with an autoencoder.
Where the images needed for the large training dataset come from is not known. However, Meta has previously indicated that it uses posts on its social media platforms for training AI models, such as integrating images into the Emu model.
More functionality in 2024
Meta announced that it will introduce more generative AI capabilities in 2024, focused on text generation and image processing. It is also testing an AI feature that helps research offers on the Facebook Marketplace and find them.