Here is one reference with autistic kids in which it was found that video modeling was better in some ways.
http://vsmproject.pbworks.com/f/video%2Bmodel%2Bvs%2Bin%2Bvivo%2Bmodel.pdf
I suspect the differences have to do with the implied judgments in "live" interactions. It turns out (there is research to back it up) people do not feel that they are being judged when interacting with mechanical objects.
People report higher levels of deviant behavior to a computer than they do to a live interviewer. Some years ago I labeled it as the "Dear John Effect." It is easier to return his engagement ring in a letter than in person for all the obvious reasons.
During live demonstrations I suspect (testing needed here) that subjects are distracted by the models more than when it is a video tape.
On the other hand, the differences may not be sufficient to require a change in one's approach. When feedback is not needed tapes are probably better. However, questions can be addressed immediately in live demos. So there is a trade off. Depending on the topic, sometimes I do not want questions until the demo is over, some times questions are preferred during the demo.
See also "feedforward learning" for some interesting ideas that would apply to teaching sports.
http://en.wikipedia.org/wiki/Feedforward,_Behavioral_and_Cognitive_Science