Pages

New Deep Learning Tech Can Make Videos Out of Still Images

Machine learning can be used for all kinds of things, and sinister applications such as Deepfakes have made people somewhat wary of this new form of technology. With all of that having been said and now out of the way, it is important to note that deep learning methods can be useful in a variety of other applications as well, such as what MyHeritage is doing by turning still pictures of relatives into moving images that can bring them to life and help you reminisce with a bit more nostalgia.

A newer application for machine learning that was developed by researchers at the University of Washington is doing something similar, except that it is turning videos of water, smoke and other types of slow flowing phenomena into full fledged videos. The videos do have a bit of a “deepfake” quality to them in that there is always something slightly off about the movements that make them somewhat easy to discern if you have a keen eye, but the fact remains that the videos still look relatively realistic and someone that isn’t looking for a fake might not be able to realize that they are looking at one.


The researchers basically fed a ton of data that showed the motion and flow of water into the machine learning algorithm, and the tech then uses this data to turn any still image of water and the like that you put into it into some kind of moving picture. The results are quite startlingly close to reality, and while the tech still has a ways to go before it can end up becoming truly worthwhile, it is still an impressive application of a technology that seems to be limitless with regards to the impact that it can end up having on the world around us.

Read next: Recreation of text styles present in any image can be possible, thanks to Facebook's new AI project

No comments:

Post a Comment