[zalas]

Here are the three original images downsampled to 720x480 using a 3-lobed Lanczos filter and then upsampled to 1280x720 using the same:

http://img444.imageshack.us/img444/7...sized01ul2.png
http://img444.imageshack.us/img444/7...sized02ox8.png
http://img444.imageshack.us/img444/7...sized03wu3.png

The only real differences I see are in the credits. And yes, I used Photoshop last time because it was convenient. These downsamples and upsamples are done with in-house software for image manipulation that my graphics research lab maintains.

Originally Posted by Mentar

Are we agreed that generally we have the 4 quality levels I outlined above?

1) Mastering HD - Airing HD
2) Mastering HD - Airing SD
3) Mastering SD - Airing HD
4) Mastering SD - Airing SD

I'll assume you mean Airing and Capturing as opposed to simply airing. 1) obviously gives the highest image quality and 4) obviously would give the lowest quality. However, 2) and 3) would give the same theoretical image quality as 4) as long as:
1. you capture losslessly (lossy captures are obviously the norm, but the lossless image quality is the upper bound) and your resampling is done ideally (most resampling is non-ideal, but a Lanczos filter gives very close results)
2. the station doesn't add any HD stuff like sponsor logos (which often change between episodes and between stations)

You see sometimes that 3) is better looking than 4), and that is primarily due to the fact that most cappers in Japan don't know how to encode properly and that fixed macroblock sizes obviously favor larger image sizes. Any differences between 2), 3) and 4) are miniscule compared to the difference between 1) and the rest.