2023-10-18 10:06:34
Yandex presented the Yandex AI Rendering Technology (YandexART) neural network, which creates images and animation based on user text requests. It generates pictures and videos using the cascade diffusion method: first it generates images and frames in accordance with the request, and then gradually increases their resolution, adding details.
Yandex assured that, compared to the previous generation of the diffusion model, YandexART creates more realistic and detailed graphics. According to the results of an internal assessment, the results of the new neural network were better in 60% of cases, according to the company’s press release (available to Kommersant).
To improve the quality of the neural network, the size of the training data set was increased by 1.5 times, to 330 million images with text descriptions. The images used to train YandexART are selected using three filtering modes. It is claimed that the new text recognition algorithm helps it better understand the wishes of users.
In addition, the neural network “develops better features of faces, eyes and hands, which allows it to make photorealistic portraits,” Yandex added. To improve the quality of generation, “reinforcement training” is used: specialists look at images of a particular object created using YandexART and select successful and unsuccessful ones. Thanks to this, the neural network better understands how to correctly depict animals, characters, objects, phenomena, etc.
In September, Yandex began testing a new way to compose “quick replies” that allow you to receive information from the network without going to the site – using YandexGPT. The company claims that it is “in contact with content publishers” and continues to refer to primary sources. But experts said there are risks for sites whose information is used by YandexGPT.
Details are in the Kommersant article “YandexGPT will retell everything.”
#Yandex #presented #YandexART #neural #network #Kommersant