Sber has presented the beta version of the new generation Kandinsky 4.0 Video neural network. The service allows you to create realistic videos based on a text description or a starting frame. This was reported by the company's press service.
The new model is capable of generating a video sequence up to 12 seconds long in HD resolution (1280 x 720 pixels) based on any text description or arbitrary starting frame. Users will be able to create videos with different aspect ratios.
The most important distinguishing features of the new model are improved visual quality — high contrast and sharpness of frames, building the overall composition of the scene, and the realism of the movements of the generated objects. This quality was achieved through the unique collaboration of scientific and engineering teams, who worked together both on the development of the new model's architecture and on the collection and filtering of data for training.
In addition, the Kandinsky team presented a fast version called Kandinsky 4.0 Video Flash, which generates a video sequence up to 12 seconds long in 480p resolution (720 x 480 pixels) in just 15 seconds.
The first users of Kandinsky 4.0 Video will be artists, filmmakers and designers. It is expected that the neural network will be available to everyone in early 2025.
Earlier www1.ru reported that the Kandinsky neural network was taught to create videos based on text.
Read materials on the topic:
Sber expands access to the GigaChat chatbot in Telegram for all users
From concept to video in a few minutes: Sber presented the beta version of Kandinsky Video 1.1