LG AI Research: The Technical Principles of the EXAONE Multimodal Model

While existing Vision AI focused on image recognition, recent trends are evolving toward generative models such as Text-to-Image, which generates images from text. OpenAI's DALL-E is a representative example, but LG AI Research's EXAONE introduces technology that goes a step further.

The EXAONE multimodal model is a bi-directional learning model that goes beyond unidirectional learning, capable of generating text from images or generating images from text. Through this, it can visualize not just simple object descriptions but also abstract concepts such as 'the energy of spring.'

One of the model's core technologies is AugVAE (Feature-Augmented Variational AutoEncoder). This is an improved model built on the existing VQ-VAE, with the following characteristics.

• Autoencoder: Consists of an Encoder that compresses data and a Decoder that reconstructs it, representing data in a Latent Space.
• VAE (Variational Autoencoder): Unlike a typical Autoencoder, it systematically organizes the latent space, allowing it to better learn the features of data and improve generation performance.

LG AI Research secured a total of 250 million image-text pair data points and used them to train the model.