Quote:
Originally Posted by blob2500
The problem was encoder-side (too), In MPEG matrix quantization internal settings.
( view here: http://websvn.xvid.org/cvs/viewvc.cg...rix.c?view=log )
If you encode with 1.3.6 and you set a MPEG matrix (custom or standard), the inter-matrix values in encoded file are wrongly changed. This reflects it to decoder-side too, because decoder use another matrix values, instead correct values to decode a file ( all MPEG-4 ASP files).
If you used H.263 quantization in your encodings with 1.3.6 I think re-encoding is not necessary. If you used MPEG... I think yes because quantization is wrong.
You can analyze your 1.3.6 encoded files with very well known Avinaptic free software, if you want see their conditions (matrix primarily)  |
Thank you. I did use the standard MPEG matrix when I experienced the problem. So it was an accidental typo in source code, i.e. while changing elem>>1 to max(1,elem)>>1 to make sure the value is non-negative (?), it happened to become max(1,elem>>1). If I understand correctly, the value was 1 for elem<=1 when it should have been 0.
AVI/xvid is still handy and useful for me, as there are a few applications that use VfW (although one can load MP4/MKV e.g. via AVS/ffms2.dll, AVI is simpler, lighter, using less memory). I do still routinely use AVI/xvid e.g. when working internally like typesetting subs, and AVI/Lagarith; even when the final encoding will be in MKV or MP4 with a newer codec. I guess this is like MP3 is still commonly used for some purposes, even though there are Vorbis, AAC, etc.
@manolito
The name Celtic_Druid brings back memories... Many years ago our sever was mirroring Celtic_Druid binaries, though mainly ffdshow.
PS
1.3.5 vs 1.3.7 Quick Tests:
"MPEG" and "H.263" are tested separately. For each:
- The .pass files are identical with 1.3.5 or with 1.3.7, except for the comment-header lines.
- The final results of 2-pass encoding are bit-identical, except for the bit-stream version strings.
- Outputs seem deterministic, i.e. if you encode twice with the exact same settings, then the final results are identical.
The test conditions are old-fashioned: a 640x480 anime episode (~25 mins) was compressed to ~200MiB, with Cartoon mode, Lumi mask, Qpel. VHQ mode=4, VHQ metric=1, VHQ for B=yes, Use chroma motion=yes.