To mount already shot video is not an easy task. Because of one incorrect replica, sometimes you have to shoot everything again or invent a director's move. However, scientists have proposed a solution to this problem. Assembly functions on the part of unsuccessful duplicates, they shifted to artificial intelligence. He was taught to pronounce the necessary phrases on behalf of the actor so that the audience did not notice anything.
A team from Stanford University has developeda curious sound track editing algorithm with elementary text input. The application uses a mechanism that decodes speech movements from different pieces of a previously recorded video. They are converted into edited material with drawing of suitable movements of the lips of announcers and actors.
For a finished video, simply add new words, remove old ones, or even swap whole replicas. As a result, the plot seems unprepared to the viewer is absolutely natural.
If the actor made a mistake in pronouncing the harvestedtext, then edit the material will be simple: just enter the adjustment directly from the keyboard. Neural Rendering machine learning technology will collect the correct version of the different words or parts thereof, said in other sections of the record. At the moment, for learning artificial intelligence will need the original video length of at least 40 minutes.
Stanford University PhD Ohad FriedHe regretted that such developments would be liked mainly by bad actors. However, in his opinion, it is worth working on technology. After all, it will have a lot of practical applications in the field of creating and editing video and other content.
As the developers assure, their methodologyVideo editing is a great help for producers and editors. But it can also be used for the production of fake rollers. Experts believe that the use of such a tool should be regulated. For example, you can notify viewers about video processing using a neural network.