Artifact around mouth when characters are speaking

[Chaobunny12]

So this is a problem I've been having for a really hard time and never successfully got around to fixing it. Basically, deinterlacing will be successful but I'll have some weird remnants of it around the mouth where a character is speaking. Here are a few consecutive frames where this happens. It's a subtle difference but it's there:




I've tried playing around with sharpeners/dehaloAlpha but I haven't really gotten any satisfactory results. I actually can't really pinpoint exactly what type of filter I should use since it's kind of a weird problem.

Here's my current script, which those screenshots came from:

Code: Select all

mpeg2source("F:\D2V files\avatar1_5.d2v")
Crop(6,0,-2,0)
Lanczosresize(720,480)
TFM(order=-1, field=0, mode=5, PP=7, slow=2)
tdecimate()
DeHalo_alpha(rx=2.5, ry=2.5, darkstr=1.5, brightstr=1.8, lowsens=50, highsens=150, ss=2) 
ConvertToRGB32()

Granted, it's a really tiny thing and most of the time I don't edit with scenes with a lot of dialogue in them, but I'd really like to be able to fix this if possible. Any ideas?

[ngsilver]

This could be related to your use of crop and resize before you de-interlace. I find it's best to crop and resize after you de-interlace. Also perhaps try spline36resize rather the lanszcos as well. You could also try doing some processing in your mpeg2 import which could help clear up this issue.

[mirkosp]

That's an in-source issue.
Basically the Avatar DVDs have some of the worst mastering seen around. The weird "shift" you see is generally caused by somehow having interpolated fields in-source mixed and matched with the original fields and I don't know what else goes on that causes it, but the interpolation is part of the issue.
There's not much one can do to fix it, if anything at all. Complex luts might be able to minimize to restore the original lineart, but at the expence of heavy artifacts like halos and ringing, so I don't recommend going there.
Instead, I found that on most sources which show this issue, just using QTGMC at its slower presets (like, "Very Slow") will help out, likely due to the usage of the mvtools and the various postprocessing QTGMC does, which means you should be able to at least get a stable look.

That said, your chain and the way you took the screenshots don't help in analyzing the issue and all that, since one first gets distracted by things such the nearest-neighbour-like artifacts shown horizontally. I would greatly appreciate a png taken from avspmod directly from just the source loading (no need to use mpeg2source's pp parameter which nate suggested: the source doesn't have blocking, so you don't need to deblock anyway, but even if you needed to, there are better deblockers around).