Better extract from video frames or set of images?

[Resio]

I'm sure this is a simple question someone mult already have been answering, couldn't find anything by searching though.
I Already made some training "experiments" with video-extracted faces with average results, i just thought it could be a good idea (and simpler indeed) to grab some thousands images from the web and extract from them, as a static image is undoubtly sharper and detailed than a video frame, and that's also a way to get more variability.
For example, i search on google images "Scarlet Johannson" and download all the jpgs it founds, to the input folder, then i extract and sort/filter...

[torzdf]

Both have their benefits, and a high quality dataset is likely to be sourced from both types of images.

The downside of images is they tend to be posed, so whilst quality is normally greater, variety in pose, expression (especially blinking!) and lighting tends to be far worse.