Which model is superior?
That entirely depends on the metric you use to measure superior. I will assume that your primary interest is how much knowledge ends up sticking in your student's brain, i.e., you are measuring superior as better scores on learning outcomes.
In your scenarios, the main source of difference is the social presence of the instructor. The primary advantage of the scenario that uses a video of the instructor is that you can communicate through gesture and facial expressions. As I understand you question, this video is only used to supplement the audio; it won't carry any meaning by itself, but may be used to emphasize. This increased social presence will likely lead to higher engagement (as also observed by paul garrett's anecdotal evidence), but its effect on learning outcomes is less clear.
There is a nice meta-analysis that looks at general interfaces that either have or don't have an embodied agent (Yee et al. 2007); in other words, if it is useful to show the picture of a person, and how realistic that picture should be. The authors find a significantly larger effect of embodiment on subjective measures ("did you think you learned more?") than on objective measures ("What was your course grade?"). They also find a generally low effect of embodiment.
More close to this question, there exists a study by Kizilcec et al. that investigates almost what you are asking. The authors performed experiments on MOOC courses on Coursea (not a flipped classroom, but close enough). One version of the course videos had audio and slides only, the other had an added talking head in the corner of the screen.
In short, the authors find in a first observational study that, if a head is present, students think they need to exert less effort, and learn more, while also rating the experience as more positive when comparing to the version without a head. However, in a following experimental study, when comparing always showing a head to strategically showing a head for some (~30%-60%) videos, no difference on course grades or likelihood of taking assignments is found. The authors did find an effect on cognitive load, which seems to be influenced by the (visual vs. verbal) preference a of the individual learner. This appears consistent with the results of the meta-analysis on embodied agents for interfaces.
My opinion is that the direct effect of a video of the presenter on learning outcomes is likely too small to be of practical relevance; then again it might increase engagement just enough to get students to watch the video that would have otherwise not watched it.