Seb,
It is a bit of both. Fundamentally it is about reducing/removing frequency components outside the region of interest and depends on the pixel rate as to where the specific frequency range of interest lies.
Google "reconstruction filters" and "anti-alias filters" for more information on the fundamental requirements.
In general though the THS8200 only supports up to 2x over sampling and depends on the pixel frequency being generated. Oversampling relaxes the performance requirements of external filters
SMPTE specifies upper frequency components whereas VESA does not, which in turn implies SMPTE may need filters whereas VESA may not.
Very few VESA systems include filters since on the receiving side there are phase alignment measurements performed to adjust the sampling phase to best match the incoming signal. In a perfectly band limited (filtered) system the sample phase would be fixed in the center phase but in reality there are lots of things corrupting the signal and causing ringing/attenuation. Think of a black white black white alternating image. If filtered at exactly the cutoff frequency with a true box filter then you would get a sin wave. Sampling this at the peak and valley would then yield the correct black white etc... whilst having the minimum ringing, EMI emissions etc... The problem with VESA is that there are many different pixel frequencies so you can't design a perfect filter without a lot of cost.
For SMPTE the problem space is significantly lower and in fact each standard has its own bandwidth limiting specifications since it was not really anticipated that a source would generate more than one standard. Once you have 2 or 3 different standards at different pixel rates you then need a programmable filter to be able to meet the individual format requirements whilst not distorting or over filtering some standards.
Hope this helps :)
BR,
Steve