QTGMC Deinterlacing Script (v3.32)

[Mini-Me]

Originally Posted by Mounir

I'm trying my script again right now with MT i'm not sure why VIt called the best settings "lossless" because it's not lossless for me , take a look:

Original:
http://img195.imageshack.us/img195/5138/originalrn.jpg
QTGMC:
http://img802.imageshack.us/img802/8199/qtgmc.jpg

settings used:
MT("""QTGMC( "Slower", EdiThreads=1, TR2=1, SourceMatch=2, Lossless=2, Sharpness=0.2, EZKeepGrain=0.6 )""", threads=2,overlap=16)

As -Vit- and Didee said, a sample would be helpful to see what's going on here. I'm not even sure what kind of image comparison I'm looking at; it could be any one of these four:

1. Woven frame and final frame: Assuming this is a pure interlaced source and QTGMC is being used correctly, there's a problem here. I don't think this is the case though.
2. Bobbed field and final frame: If this is the case, it would be helpful to see the neighboring bobbed fields from the source too (or better yet, the source clip).
3. Original separated field and supposedly (but obviously not) lossless counterpart from the final frame: Assuming this is a pure interlaced source and QTGMC is being used correctly, the last smoothing is doing more damage than I'd expect.
4. Original separated field and generated pair from the final frame: This kind of comparison wouldn't really make any sense; it would be better to see the whole final frame.

A thought just came to me though: The temporal smooths in QTGMC might GREATLY benefit from "gamma-corrected" smoothing, when it comes to thin horizontal details. See here to get an idea of what I'm talking about. The basic idea is that at ordinary gammas, any kind of smoothing or averaging ends up reducing the overall luminosity of areas (the softened grill of the car has been darkened excessively). To correct for this, you would reduce the gamma, perform your operations, and increase the gamma again.

As a quick test to see if it might help, you could try sandwiching your QTGMC call between with two calls to levels or LaTo's Smoothlevels (the first reducing gamma and the second increasing it again). The optimal solution requires 10+ bit processing in Avisynth, but if the Smoothlevels solution makes the picture look better in the meantime, why not? If time is no object, you could even repair very subtle detail (least significant bits) lost in the gamma shifts by doing the following:

• Run QTGMC normally and save the result as A.
• Reduce source gamma, run QTGMC on darkened source, and increase the gamma of the result, and save the result as B.
• Run some kind of repair filter that replaces pixels from B with pixels from A if they're in the same brightness neighborhood (within 1 or 2 luma or so). I could probably think of some hacky way to do this off the top of my head, but I don't want to embarrass myself in the presence of my betters.