Meta introduced 'CM3leon,' an AI-powered image generator that can transform text-to-image and image-to-text generation.
2
Not diffusion-based
Meta's CM3leon employs the 'attention' method in transformer models instead of diffusion. This method allows parallel processing and increases processing speed.
3
Image and text
Meta's CM3leon can create sequences of texts and images. The ability to generate images and text improved its performance in various tasks.
4
Parameters
Meta's CM3leon has 7 billion parameters, outdoing OpenAI's DALL-E 2 with 3.5 billion. It was trained on millions of licensed Shutterstock images.
5
Captions
Meta's model can perform various text tasks, including generating short or long captions and answering questions about an image.