Перевод на русский готовится — пока показан английский текст.
Sber releases Kandinsky 6.0 Video AI model that generates video with sound
Коротко
Sber announced Kandinsky 6.0 Video, a generative model that creates video together with a synchronized audio track, including speech, music and ambient sound. The model is available in GigaChat and can produce clips of up to five seconds, a limit Sber attributes to computing costs and says it plans to raise. Sber says the new version conveys real-world physics more accurately than the previous one.
Читать полностью 2 мин чтения
Sber announced an updated generative model, Kandinsky 6.0 Video, which can automatically add an audio track to video, Vedomosti reported, citing a company press release. Vedomosti quoted Anton Frolov, senior vice-president and head of Sber's generative AI development unit, as saying: “We want generative AI to become as natural a tool for a creator as a text editor or a phone camera is today.” 3DNews reported that the company describes the model as the first Russian neural network able to create video with synchronized sound. [ 1 , 2 ]
According to the documents, the model generates speech, music and environmental sounds matching the content of the video. Vedomosti reported that a user only needs to set a text prompt or upload an image on which the neural network builds a clip with sound. 3DNews reported that users describe the scene in text, upload a first frame and describe the desired sound, with quality options of SD, HD and Full HD, audio generated at 44 kHz, and an option to create video without sound. The outlet added that the model can search for references itself if it lacks information about how an object in the request looks, and that it is available in GigaChat, which Habr News also reported, along with availability on Hugging Face. [ 1 , 2 , 3 ]
All three documents noted that clips with sound are limited to five seconds because of significant computing costs. Vedomosti and 3DNews said the developers plan to increase the duration over time; Habr News said the same, without giving a target or date. [ 1 , 2 , 3 ]
Sber said the new version better conveys the physics of the real world, so that the movements of people, animals and objects in frame look more natural and the scenes more plausible, according to Vedomosti and 3DNews. 3DNews said the developers estimate the model is better than the previous version in an average of 71% of cases across all criteria, and reported that Kandinsky 6.0 Video is available to developers under the MIT license, which the outlet said allows free integration into their own products; the other documents do not mention the license. [ 1 , 2 ]
Vedomosti placed the release in the context of Sber's earlier AI announcements, noting that on 10 September the company presented GigaChat 3.5 Reasoning with a reasoning mode, which it positions as the only Russian open model with such a mode. [ 1 ]
Почему это важно
3DNews reported that Sber calls Kandinsky 6.0 Video the first Russian neural network able to create video with synchronized sound, which would put a domestic model into a field dominated elsewhere. According to 3DNews, the model is available to developers under the MIT license, which the outlet said allows free integration into their own products, and Sber expects it to be used by content creators, marketers and designers. Vedomosti reported that Sber positions its AI releases as part of a broader push, after presenting GigaChat 3.5 Reasoning in September.
Ключевые факты
- Sber released an updated generative model, Kandinsky 6.0 Video, which can automatically add an audio track to video. [ 1 , 2 , 3 ]
- The model generates speech, music and environmental sounds matched to the content of the video. [ 1 , 2 , 3 ]
- A user can set a text prompt or upload an image, on the basis of which the model creates the clip. [ 1 , 2 ]
- Clips with sound are limited to five seconds because of significant computing costs; Sber says it plans to increase the length over time. [ 1 , 2 , 3 ]
- The model is available to users in GigaChat. [ 2 , 3 ]
- Sber says the new version conveys real-world physics more accurately, with the movements of people, animals and objects in frame looking more natural. [ 1 , 2 ]
- 3DNews reported video quality options of SD, HD and Full HD, and said audio is generated at 44 kHz. [ 2 ]
Подтверждено несколькими источниками
- Sber announced Kandinsky 6.0 Video, a generative model that produces video with synchronized sound. [ 1 , 2 , 3 ]
- The model can generate speech, music and ambient sounds. [ 1 , 2 , 3 ]
- The maximum length of a clip with sound is currently five seconds, with longer videos planned. [ 1 , 2 , 3 ]
- The model is available in GigaChat. [ 2 , 3 ]
Пока неясно
- Kandinsky 6.0 Video is available to developers under the MIT license, which 3DNews said allows free integration into their own products. Only 3DNews reported the license; Vedomosti and Habr News do not mention it.
- The developers estimate the new model is better than the previous version in an average of 71% of cases across all criteria. This figure appears only in 3DNews and is attributed to the developers.
- The model generates audio at 44 kHz, which 3DNews described as standard quality for many streaming services. Single-source detail from 3DNews.
- When longer videos will become possible, and how much the limit will be raised. The documents say developers plan to increase the duration but give no target or date.