Let's start with the original video, which was found on youtube:
That's definitely some good starting material. I didn't fully appreciate how difficult this background was going to be though (more on that later). I chose to focus on the first ~45 seconds of the clip, since that had a lot of upper body posing and nothing moves too much. I used the liquify effect to create the changes, just as you would to create a 2D morph in photoshop. The difference here, is that you can animate the changes made by the liquify tool itself, and you can have those changes track a feature of the video. There's some tutorials out there that show this kind of thing to give you an idea (try here or here).
So, using a tracking point on the arm, and another for the chest, I made a few early changes that make him look a little lopsided:
After doing a few experiments with mattes and masks I decided to try something more extreme that would destroy the illusion by bleeding over into the background, with the hope of correcting that later. So the next iteration has even more disparity:
Finally, I worked on getting all of the morphing/liquify elements adjusted to get something like this:
At this point the severe background distortion sort of kills it for me. This was the most frustrating part of making this video, and the results are far from perfect. I tried to use a color picking mask to select out the morphed bodybuilder which is composted over the original, un-morphed video. Here's a screen shot:
So everything shown in orange here will be replaced with the original video footage, which should remove the distortion in the background image. The problem I had was finding a stable filter that isolated just the bodybuilder, but nothing else. I'm sure that there's lots of tricks that I'm not using that could make a better result, but this was the best I could do. Finally, some of the morphs and changes were coming out too fast (especially since I wasn't very careful with rendering out sections to get a sense of speed). So I also slowed the video down a bit to exaggerate the movements. Which brings us to the final video:
