Beijing AI Academy Unveils Groundbreaking Multimodal Model, Challenging Industry Giants with its Technological Leap

The government-supported Beijing Academy of Artificial Intelligence has unveiled a new model, which it claims can surpass Stable Diffusion XL in certain tasks, marking it as their 'biggest technological contribution'.

Chinese AI companies are striving to circumvent the limited availability of high-tech chips and less accessible funding compared to the US. They are working hard to keep pace with the swift progress of industry pioneers like OpenAI and Google. BAAI, a non-profit organization, aids in boosting the growth of China's AI sector.

BAAI's newest version of the Emu3 multimodal model employs a straightforward architecture to train models to interpret images and generate video segments, as announced at a meeting in Beijing on Monday. Unlike conventional models that only manage one form of data, multimodal models are designed to comprehend various kinds of input data, including text, video, and audio.

The leader of BAAI, also referred to as the Zhiyuan Institute, Wang Zhongyuan, stated that the latest model represents the most significant technological advancement in recent years from the organization, which is six years old.

Emu3 utilizes a consolidated AI framework that converts text, pictures, and video snippets into a jumble of tokens. These tokens are used to initially train a singular model. A token represents the tiniest data segment, like words, image fragments, or video scenes, that an AI model can handle.

This method eliminates the necessity of merging various task-specific models to manage different data types, simplifying and enhancing the efficiency of training a multifaceted AI model.

According to BAAI, Emu3 surpasses several recognized targeted-task models, including the Stable Diffusion XL used for image creation, and the multimodal model LLaVA, in terms of comprehending and generating images.