With the processing power that is able to be tapped from not only a computers APU but also GPU’s, machine learning is now able to perform the sort of compute that could only be dreamed about before. One simple example is how the facial features of one person can be perfectly superimposed over another as is the case in the following where Arnold Schwarzenegger’s face is perfectly transitioned and superimposed over Bill Hader’s (and likewise, transition seamlessly back to his own face).
The ramifications of this is quite clear where the compute power along with code now exists where fiction can meet non-fiction head on (and you don’t need very expensive/specialized equipment to accomplish it; even smartphones have the processing power to handle this). The above was based on code (DeepFaceLab) released publicly on Github. Even in Star Wars: Rogue One, the process and technology wasn’t even close to this level when it came to recreating Tarkin and Leia’s facial features on the original actors face. Moral issues will of course be crossed (both the Cushing and Fisher family gave their ok) because it will actually be possible to recreate and even do new works using AI and machine learning to create younger versions of a actor/actress utilizing archived imagery for the face sets. The point of course is that there will be a time where even video will no longer be admissible in a court of law due to how it can be easily doctored (as the technology and hardware processing gets even better).
Similarly, machine learning algorithms were used to create a higher resolution version of Diablo II’s opening cinematic by interpolating and adding details that allows each frame to be displayed at a higher pixel resolution than the original. Similarly, such code can be used to increase the detail of existing textures without having a graphic artist painstakingly doing it from scratch.
The technology isn’t to the point where it can be done in real time; the point is that the technology exists where it can assist in remastering textures and other graphical assets (mainly 3D/vector based artwork). In the case of sprite based games like Diablo I and II, this technology could really only be used to enhance the cinematics and to recreate better textures for creating higher quality sprites without having to try to recreate the gameplay in 3D (where the challenge is not losing the feel of the original).
Vectorizing pixel based art (depixelization and upscaling) is something that was researched and done many years ago though. Utilizing machine learning and AI into the mix would be yet another tool to aid in remastering graphical assets without busting the production budget in the process. It remains to be seen whether or not Blizzard intends to actually offer up a remastered version of Diablo II (the point of this entry is to note there are more tools now at these companies disposal to accomplish some of the work).
UPDATE: Someone else “remastered” the intro cinematic for Diablo II: Lord of Destruction utilizing machine learning. The limitations can be seen in high motion scenes though where the already poorly compressed (quality wise) source has a lot of pixelization (so there simply is not enough data to interpolate and upscale except to upscale the pixelization itself).
