Yandex presented the updated neural network YandexART 1.3

The company more than doubled the data set on which the model was trained. And latent diffusion technology requires less computing power while allowing more advanced graphics to be generated, Yandex explained.

How does latent diffusion technology work?

Latent diffusion technology creates an intermediate representation of an image in a latent code format. This is a compact description with basic information about the image in a compressed format. After this, the neural network in a single step expands the code into a full high-resolution image. As the company notes, this method is more efficient than multi-stage image refinement using cascade diffusion.

Yandex also increased the data set on which the neural network was trained by two and a half times. Synthetic tests have been added. The data set itself was expanded to over 850 million pairs of images with text.

In order for YandexART to take into account more details of industrial data, the new model uses not one, but two text encoders, with the help of which the model translates a text request.

In addition, now with YandexART you can create images in different formats: 16:9, 4:3 or 3:4. The update is already live in Masterpiece and will soon appear in other company services.

The company first introduced the YandexART broadcast neural network in 2023. Yandex noted that the service “understands the Russian cultural code” and knows important symbols, characters and personalities from Russian history.