Caution: loooooong post !!!
By now it is quite widely scientifically accepted that the mind works as a massively parallel hierarchical pattern processing unit
I recommend reading a book called “How to create a mind” by Raymond Kurzweil, it is really interesting (he is a lucid dreamer, too)
In short: every information in the mind is treated as a series of hierarchically organized patterns, e.g. areas of different brightness are organized into edges, which become lines and curves, which in turn are organized into letters, letters into words, which are connected to the stored meaning(s), then words are organized into sentences.
The patterns which get activated the most get reinforced by stronger neural connections
The mind also continually makes predictions on the next patterns it will recognized, based on the pattern already recognized so far
This is why the famous trick of properly reading sentences containing words having only the correct starting and ending letter works just fine
Note: modern AI learning algorithms, e.g. computer image/speech recognition work using the same approach
The same happens with faces, edges are interpreted as eyes, mouth, nose, etc, with some specific characteristics, and stored as patterns; next step is storing the pattern of those peculiar face characteristics together with its “label” (name of the person), which in turn is a pattern of symbols (letters)
Interestingly enough, faces are the first thing the mind uses to train its visual area, by studying and storing mommy’s facial features
As children we are not born blind, our eyes work, just our minds don’t have stored patterns that we can use to build meaning from the mess of visual stimuli we receive; we start “seeing” things as soon as we match enough stored patterns
Since we are so well trained in recognizing faces, this is why it’s so easy recognizing faces even when they are not there (e.g. Moon and Mars craters)
Now, is the mind capable of creating something “new” by assembling stuff it already knows ? I’d say there is no doubt the answer is yes, you do it every time you talk, write, cook, drive, whatever.
Is the mind capable of creating a believable mental image starting from stored visual and content patterns ? Most definitely yes, you experience this when it creates every part of the dream environment, e.g. buildings having a reasonable amount of bricks and windows, lifelike roads with streetlamps, fences, houses, trees, road signs etc.
Is the mind capable of building a never mate face from the stored patterns ? Based on the above, and knowing the mind is particularly well trained in recognizing faces, I’d say we can safely answer our mind can do it easily.
Now, having to create a face in the dreamscape, would our mind prefer re-using a face only seen once (a never met before supermarket cashier), or build it from scratch, assembling known patterns ?
I’d say it depends: if the cashier has in some way aroused our interest, even subconsciously (e.g. she was cute) perhaps its neural connection is strong enough that our mind can recover it and reuse it (maybe, at least partially)
On the other hand, if the cashier was all-around average, most likely its connection will be over-ridden and reused (re-wired); at that point, I guess our mind would be more likely to just make the new face up
TL;DR: IMHO during dreams the mind does both, it creates new faces, but also reuses known ones