YandexART

2024

YandexART 1.3 with latent diffusion

Yandex on April 24, 2024 introduced the YandexART 1.3 diffusion neural network. She switched to a new technology for generating images - latent diffusion. In addition, the datacet at which the model was trained was increased 2.5 times. Thanks to this, YandexART understands text queries better and creates even more realistic images in different formats. The updated version of the neural network is already being used in Masterpiece, and later it will appear in other Yandex services.

Latent diffusion technology consumes less computing resources and allows for more realistic graphics. It forms an intermediate representation of the picture in the form of a latent code - a compact description containing the main information about the image in compressed form. Then the neural network expands the code into a full-fledged high-resolution image in one step. This approach is more effective than multistage picture refinement in cascade diffusion.

source = Yandex

To make the model better understand user requests, synthetic texts - detailed descriptions of images generated by the neural network - were added to the datacet on which it was trained. The dataset itself has been expanded to more than 850 million picture pairs with text. In order for YandexART to take into account more details from the promt, the new model uses not one, but two text encoders. With their help, YandexART recognizes the user's text request and translates it into machine language.

In addition, the YandexART update will give users the ability to create images in different formats, such as 16:9, 4:3 or 3:4. Now the images created in the Masterpiece can be used, for example, for magazine covers or photo reviews on the phone.

According to measurements by Yandex assessors, the YandexART 1.3 neural network wins 57 percent of the time compared to Midjourney V5.2 and 63 percent of the time compared to the previous version of YandexART 1.2.

Availability in Yandex.Directive

Entrepreneurs who do not have a site will be able to advertise their goods and services in Yandex.Directive using neural networks. Yandex announced this on April 24, 2024. The YandexART neural network will draw unique illustrations. Read more here.

Access APIs

On April 9, 2024, Yandex"" announced the opening of access API neuronets to YandexART illustrations. Companies can use the tool in the Yandex Foundation Models service on the cloud platform. Yandex Cloud The neural network is available for free.

Using the YandexART neural network API, entrepreneurs can create images for advertising and social networks, design layouts and illustrations for social networks and books. In addition, the API will help build picture generation into services and web applications.

source = Yandex

It is noted that several companies have already tried YandexART as part of closed testing. So, one of the major retailers (his name is not called) uses a neural network to create a unique design of gift cards. As noted in Yandex, using this artificial intelligence, customers will be able to enter a text request, after which they will be able to choose from several generated options.

From April 9, 2024, access to the service is open to more users. YandexART is based on several of Yandex's own products. The company notes that the YandexART neural network is constantly being finalized to improve the quality of generation. For example, the company's specialists look at several images and note the most successful and appropriate requests. This helps to better distinguish objects and speed up work with the neural network, which is especially important for business scenarios, Yandex emphasized.

The service is based on the cascade diffusion method used to create an image. First, the program generates pictures and personnel in accordance with the user's request, and then it gradually increases their resolution, saturating them with details. Thanks to this, realistic and detailed graphics are achieved, added to Yandex.

2023

Ability to create images of goods on Yandex.Market

Sellers Yandex.Market can now create images of goods for free using diffusion neuronets YandexART. It is enough to upload photos of products on a solid background - and the neural network will generate realistic images that will show the product as best as possible in a suitable environment. For example, if it is a coffee machine, the neural network will add it to the interior of the kitchen, and if the toy - to the setting of the children's room. This was Yandex announced on December 25, 2023.

This technology not only simplifies the creation of content, but also reduces the cost of photography. The market calculated that sellers spend from 5,000 to 7,000 rubles on content for one product card. Now you can do without professional filming: for example, it is enough to take a smartphone with a good camera, photograph the goods against a neutral background, and everything else will do the neural network.

You can create images of goods in the seller's personal account on the Market. If the resulting options do not suit you, the service allows you to generate new ones as many times as you like. The best images can be immediately added to the product card on the Market by pressing one button. The seller can also see pictures in good quality for free - for example, to use on other sites.

As of December 2023, the new tool works in beta and is available for some categories that include more than 600 thousand products. The market will develop the technology and add new features to it.

Recently, sellers on the service also have the opportunity to use the neural network to generate names and descriptions of goods, as well as advertising banners, for free. In November 2023, Market said that in this way already 12,500 sellers came up with the names of goods, 17,700 - at least once created a product description.

Product Announcement

On October 18, 2023, Yandex introduced the YandexART neural network, which creates realistic pictures and animations at the text requests of users. According to the developers, the technology "knows" famous places and cities of Russia, outstanding personalities of different eras and characters of cartoons and fairy tales. It is already integrated into the Masterpiece application.

Compared to the previous generation diffusion model used in "Masterpiece," YandexART creates more realistic and detailed graphics. According to the results of the internal assessment of assessors, its results are better in 60% of cases, according to Yandex.

To improve the quality of the neural network, the size of the training set of data was increased by 1.5 times, to 330 million pictures with a text description. Images used for YandexART training are sampled using three filtering modes.

With YandexART, you can create stunning images: beautiful, exciting, funny. Even if you do not know how to draw at all, you just need to write a detailed and understandable request to neural networks, the company said.

As the developers explained, YandexART creates images and animation by cascading diffusion - first generates pictures and personnel on demand, and then gradually adds details and increases their resolution. It is noted that artists, musicians, designers and illustrators can use YandexART for creative experiments and search for inspiration. The neural network helps them create album covers, theatrical scenery, costumes and other art objects.

In addition, YandexART creates images for advertising. Yandex users . Businesses "can choose for the announcement the picture generated by the neural network. In the future, the developers plan to add a neural network to the "Keyboard" (users will be able to generate a picture and set it in the background) and to other services.^[1]

Notes