Meta recently announced an artificial intelligence model that can generate both images and text content.CM3Leon, claiming that it only requires one-fifth of the training data required by other similar models to achieve the same effect.
CM3Leon itself is a general-purpose artificial intelligence model that can handle different tasks simultaneously, including generating images through text descriptions, editing images through text descriptions, and automatically generating image text descriptions.
Its inherent causal and hidden hybrid model architecture allows it to ignore or conceal specific content during training, thereby producing results that better align with user expectations. Furthermore, CM3Leon was trained using only 30 billion text parameters, compared to OpenFlamingo, which uses up to 400 billion parameters, and Flamingo, which uses up to 1000 billion parameters, achieving nearly identical performance.
However, Meta has not disclosed whether it plans to open CM3Leon to the public, and it may only be used as an internal application tool.
