The updated YandexART 1.3 allows users to create images in various formats, including 16:9, 4:3 or 3:4.
Yandex announced YandexART 1.3, an enhanced version of its artificial neural network that converts users’ textual descriptions into ready-to-use images and animations. YandexART 1.3 has improved its technology for generating images through latent diffusion. Additionally, YandexART increased the image dataset used to train the model by a factor of 2.5.
These improvements allow YandexART to better understand text commands and create more realistic images in different formats. The new version of YandexART is now available in the Shedevrum application worldwide.
The hidden diffusion technique requires fewer computational resources and facilitates the creation of higher quality graphics. The process starts by developing an intermediate image representation known as secret code. This is a compact description that contains basic information about the image in compressed form. The neural network then converts this code into a high-resolution image in a single step. This technique is more efficient than the multi-stage refinement of gradual diffusion.
Yandex also added detailed image descriptions, known as synthetic texts, to the training dataset, which the neural network generated, to enable the model to better understand user prompts. Yandex expanded the dataset to include more than 850 million image-text pairs. Additionally, Yandex included two text encoders in the model to enable YandexART to take into account more details from user prompts. These encoders enable YandexART to accurately interpret text commands and convert them into machine-readable data.
The updated YandexART allows users to create images in a variety of formats, including 16:9, 4:3 or 3:4, making them usable on magazine covers, television and more.
Internal evaluations show that YandexART 1.3 outperformed Midjourney V5.2 in 57% of trials and YandexART 1.2 in 63% of trials.
