No, pixel sizes do not need to match. As your data is not georeferenced anyway, the pixels won't line up after referencing and will be more or less distorted anyway.
But: The better your reference data, the better you'll be able to place your control points!
By "geometric model" I assume you mean which model to use for interpolating between the control points? This is not easily answered. ERDAS Imagine should give you some good pointers in its documentation, already.
A simple affine reference is easy and will not distort your source data significantly, but it will have trouble covering many variations that are common in airborne data, especially old one.
Strong variation in airborne data can be caused by movement of the sensor (shaky plane/platform) and vertical variation in terrain. If you have hilly terrain instead of flat, you will need a lot more sample points and distortion to stretch the distorted aerial imagery into place. In these cases, you'll absolutely not be satisfied with an affine model and will need to go for a polynomial one.
It also depends on what accuracy you need for your referencing. Do you intend to do spectral analysis (where you want sub-pixel accuracy) or are you just mapping it roughly for visual reference?
If you have questions regarding specific models and their suitability, expand your question.
All of these questions have an impact on what approach is recommendable to georeference data. Georeferencing will always require human creative input and thus is a bit of an art form (mostly for placing control points), but there's lots of science behind it to put you on the right track.
One last note on polynomial models: The higher polynomial degree, the more it will distort your data, and it will reduce your statistical errors (RMSE). However, it will not necessarily improve your quality of georeferencing, and can even be vastly worse than a polynomial correction of lower degree. Also, high degrees have a vastly higher computational cost.