Several models are particularly popular today. Firstly, there are autoencoder models, including variational autoencoders (VAE)
. They learn based on data structure and can generate new examples while preserving the essential characteristics of the original data set. VAEs, for example, are used in the field of computer vision to create new images or restore original images from noisy data.
Secondly, we see widespread use of generative adversarial networks (GAN)
, which consist of two networks: a generator and a discriminator. The generator creates new data, and the discriminator attempts to determine whether they are real or artificially generated. As a result of this "competition," models can create high-quality images, text, and other content. For example, GANs are used to create realistic images of people's faces that do not actually exist (the "This Person Does Not Exist
Another important player in this field is the Transformer, specifically the GPT (Generative Pretrained Transformer) model developed by OpenAI
. This model uses the Transformer architecture to generate coherent and quality text based on previous context. Uses include text translation, question answering, automatic article writing, and even the creation of new stories or poems.
All these models are used in different ways in the real world, and they all represent significant steps forward in the development of generative AI.