Sberbank announced a test launch of a new version of the Kandinsky neural network for video generation. According to the developers, the proprietary algorithm has become more accurate in understanding requests and creating more realistic short videos, and they are created even faster than before.
Kandinsky 4.1 Video generates a video sequence up to 10 seconds long based on a text description or the original frame. The video resolution is SD (720×576) or HD (1280×720). Sberbank representatives noted that the quality of the material was improved with the help of additional training (Supervised Fine-Tuning, SFT) on a specially selected dataset: the images were selected by designers, photographers and artists with specialized education.
“…The model has become much better in all respects: in compliance with the prompt, visual quality, quality of motion generation, as well as the ability to model the physics of the world,” said Andrey Belevtsev, head of the Technological Development block and senior vice president of Sberbank.
More cinematic videos increased the hardware requirements, but after using distillation and acceleration methods, developers were able to increase the generation speed by more than three times compared to the original. The quality did not get worse, and in some scenarios it even improved.
The Kandinsky 4.1 Video neural network is already available to GigaConf conference participants, as well as some designers and artists. The algorithm will become publicly available later, but the exact dates have not yet been announced.