Sber has launched Kandinsky 6.0 Video, an AI model integrated into GigaChat that generates Full HD video clips accompanied by synchronized speech, music, and ambient sound. This marks the first Russian-developed neural network capable of producing video with integrated audio.
To generate content, users provide a text prompt, an initial reference frame, and a description of the desired audio. Currently, the system supports clips up to five seconds in length, with plans to extend this duration in future updates. Users can select output quality from SD, HD, or Full HD resolutions.
The model produces audio at a 44 kHz sample rate, matching standard streaming quality, though users may also opt to generate silent videos. If the system lacks specific visual data for an object, it can autonomously search for references to improve accuracy.
Developers claim the new model demonstrates improved physical realism, outperforming the previous version in 71% of test cases. Kandinsky 6.0 Video is released under the MIT license, allowing for free integration into third-party software products.
