Yandex releases YandexART 1.3: next-level realism and improved user prompt understanding

Yandex recently announced the release of YandexART 1.3, an upgraded version of its neural network that turns text descriptions from users into ready-to-go imagery and animation. YandexART 1.3 features improved technology for generating images via latent diffusion. Moreover, the image dataset used for training the model has grown 2.5 times. These enhancements enable YandexART to better understand text prompts and create more realistic images in different formats. The new version of YandexART is currently available worldwide in the Shedevrum app.

Latent diffusion requires fewer computational resources and facilitates the creation of higher-quality graphics. The process begins by developing an intermediate image representation known as latent code — a compact description with essential information about the image in compressed form. The neural network then transforms the code into a high-resolution image in a single step. This technique is more effective than the multi-stage refinement of cascade diffusion.

To improve the model's understanding of user prompts, Yandex added detailed image descriptions generated by a neural network — known as synthetic texts — to the training dataset. The dataset itself was expanded to include over 850 million image-text pairs. Furthermore, to make YandexART consider more details from user prompts, two text encoders were incorporated into the model. These encoders enable YandexART to accurately interpret text prompts and convert them into machine-readable data.

The updated YandexART allows users to generate images in several formats, including 16:9, 4:3, or 3:4, making these images usable as magazine covers, homescreens, and more.

Internal assessments indicate that YandexART 1.3 outperforms Midjourney V5.2 in 57% of cases and YandexART 1.2 in 63% of cases.

About YandexART

Yandex AI Rendering Technology (YandexART) generates imagery and animation through cascade diffusion. The neural network does that from text prompts provided by users, gradually improving image resolution and adding finer details. Trained on a database of 850 million pictures with descriptions, YandexART employs a unique text-recognition algorithm to interpret user queries with contextual precision and cultural relevance. The network is capable of generating intricate patterns and adopting specific visual styles, including photorealistic portraits.