Okay, it was a lot more complex than it looks. The drop in resolution completely hides the handoff; in the high-res version, you can see that there's actual overlap where both hands are on the can simultaneously. I didn't get it perfect, but with more time I actually could have.
The can is on a plate that's painted fluorescent green for keying out. I did it that way specifically to test the greenscreen capabilities of the camera and the software. The results are mixed; greenscreen is absurdly difficult, and this was more difficult than most because of the lighting needs.
I munged up the lighting a bit. I had a 1K open face bounced off of the ceiling for a fill light and a 1k fresnel through a diffuser for a key, plus one of the chinaballs above my head on screen left. The big bounce light ended up getting reflected off the back of the couch in a really inopportune place, which made matching that particular piece all the more difficult, particularly with shadows.
Another downside is the combination of being slightly out of focus and the 4:2:0 color sampling of the camera, and the edges of the green plate weren't well-defined, which made pulling a key really hard: the fringe was nearly impossible to get rid of without screwing up the rest of the image. It's pretty obvious in the high-def version. I do think that a normal greenscreen with separation and dedicated lighting has a good chance of working, but it's still gonna be touchy.
I also muffed the timing on the handoff. I should have reshot the first one (the one without the hat). And I could have played with the splitscreen bezier keyframes to get a smoother mix, and I could have hand-painted the critical frames to get the shadows right and smooth the alignment of the keyed-in portion, but again, it's a test, a bit of learning that I wanted to do. In that, it succeeded.
Big lesson is that high def is a lot more unforgiving, and things like focus and lighting are critical.