“Yandex” released an updated neural network YandexART 1.3

Yandex has unveiled an updated YandexART 1.3 neural network with improved capabilities. Thanks to new image generation technology and a 2.5 times larger dataset, the artificial intelligence now better understands queries and creates more realistic pictures.

YandexART 1.3 with improved capabilities.

The latent diffusion technology not only improves graphics, but also requires less computational resources. It creates an intermediate representation of an image in the form of latent code, which the neural network unfolds into a full-fledged image in a single step. With YandexART 1.3, you can create images in various formats: 16:9, 4:3 or 3:4.

To better understand queries, the developers have added synthetic texts to the dataset — detailed descriptions of images generated by a neural network. In addition, the new model uses two text encoders instead of one. With their help, YandexART recognizes the user’s text query and translates it into machine language.

The new YandexART 1.3 is already available in Masterpiece and other Yandex services.