Research from the UC Berkley shows that AI-dri­ven image-manip­u­la­tion may be far more advanced than you might believe:

This paper presents a sim­ple method for “do as I do” motion trans­fer: giv­en a source video of a per­son danc­ing we can trans­fer that per­for­mance to a nov­el (ama­teur) tar­get after only a few min­utes of the tar­get sub­ject per­form­ing stan­dard moves. We pose this prob­lem as a per-frame image-to-image trans­la­tion with spa­tio-tem­po­ral smooth­ing. Using pose detec­tions as an inter­me­di­ate rep­re­sen­ta­tion between source and tar­get, we learn a map­ping from pose images to a tar­get sub­jec­t’s appear­ance. We adapt this set­up for tem­po­ral­ly coher­ent video gen­er­a­tion includ­ing real­is­tic face syn­the­sis.

Every­body Dance Now

Previous ArticleNext Article

Leave a Reply

Your email address will not be published. Required fields are marked *