as a result, sberbank selected videos for datasat, and then generated a description for them, again withp of nerosey... that’s what they trained, in general, this is a global problem, the sora project, the famous open ai, was called a real breakthrough, it was presented in february this year and became the first neural network capable of generating very high-quality high-resolution videos up to a minute long, they can contain several characters, different in the complexity of the movement of objects, and at the same time amazing detail, before that, google lumier was considered one of the most promising projects, the videos there are short, but very realistic. google came up with the idea of ​​teaching a model not just to draw individual frames, but to understand how objects are located in space and how something changes over time. but the kandinsky video is available for free, and as for longer videos and cinematic quality, how is it that the developers in their article on habré called it their next goal: well, probably and from yandex masterpiece we should also expect an update to the