You could do it using breath made from smoke, but it might be challenging matching up all of the elements. Not simply in terms of footage, but in terms of performance and nuance of your actors.
Lock Your Picture
You don't want to be sinking in effort until you are certain of which frames you will be performing surgery on. Once your edit is locked, drop the raw frames into a 16 bit half EXR per shot. Assuming you have shot this on a DSLR or like camera, isolate your final cut from the VSE and render out.
If you have several to many shots, maintain solid control over your organization. One directory per shot, with versions as subdirectories or an alternative approach that works for you. Bear in mind that if you are performing a dissolve or an over where two shows are displayed at the same time, you will need to make sure that you have a full series of the entire frame count, including the overlapping portion, in your half EXRs.
Gather the Plates
Ideally you would have a library of plate shots, shot against black. Consider a back light to bring just enough lift into the breath to give distance between the lower black of your background and your foreground breath. You could capture these using a DSLR without much effort, assuming you can find a cold enough area to work in for a bit. Gather up many plates so that you can combine them in ways that suit the breath required in your shot. If in doubt, capture more rather than less.
If shooting is unrealistic, consider rendering out plate elements with variations using your smoke concept or gathering up found footage that might be available online with permission to use.
Use associated alpha (sometimes sarcastically referred to as the One True Alpha) on the plates based off of a luminance key pull, making sure to bring up your lower end enough to threshold out the black you are photographing against. If grading is required, make sure to unpremult the alpha otherwise you will incorrectly skew the RGB to alpha relationship leading to glowing or darkened fringes.
Depending on the nature of the breath element, you might be able to get away with a 2D manually tracked element with an empty and some scaling. However, if you have a shot that warrants it, consider tracking it and laying in your elements as it shouldn't be too complex and potentially add tremendously to your work. You gain the ability to add subtle depth of field and possibly some depth cuing for free.
If you do track your shot, consider adding a subdivide or two to your floating plates and subtly warping them with a quick sculpt or curve. This will add some perceptual parallax as the camera moves and more greatly sell the effect.
Matching Footage to the 3D View
I'm not sure how to see the video so I can position exactly frame by
frame a simulation either.
From the Properties panel (hotkey N) you should see "Background Image". Set accordingly. To keep your sanity, position the camera at a perpendicular position to the grid and position your elements accordingly. If you are going to attempt a 3D effect, make sure to set your camera lens to roughly what your shot was conducted on, and set the sensor size to APS-C or whatever gate size you shot on. If you are going to do a strictly 2D approach, this isn't an issue.
If you are attempting something strictly 2D, where you are using an orthogonal camera and floating 2D planes,this can be a little tricky if you aren't careful. First, make sure that your render dimensions match the dimensions of your footage. Second, if you are seeking to perfectly match a plane to the viewport, you will need to set the scale of your pla such that it matches the ratio of your frame. That is, for 1920x1080 footage, your scale would be 1.920 by 1.080. If you don't use the fractional values, you will end up triggering some cubic scaling on your images and the result will end up a blurry soup, so be aware.
If plausible, try the 3D approach for fun.
Here is one approach, loosely summarized above, from a larger budget production: