Saturday, February 13, 2021

Technology Continues to March On

In previous posts, I've written about advances in technology that have enabled Orphan, and other fansub teams, to get more out of analog sources. These included:

  • The Domesday Duplicator, which taps into a laserdisc player's laser and records the raw RF data, bypassing the player's aging analog electronics. The RF information is then processed in software.
  • An uncompressed VHS capture setup, which eliminates compression artifacts from VHS capture cards and allows software to deal with VHS issues.

These improvements have produced noticeably better raws from analog sources and, as a result, a raft of re-issues of old projects.

Technology does not stand still, though, and improvements happen almost continuously, particularly in software processing. For example, ld-decode, the software for the Domesday Duplicator, is now on its sixth revision and has improved in every aspect. A continuous "diet" of rips from flawed laserdiscs has made the software much more robust against bit rot and other physical problems. The latest development allows direct decoding to YUV colorspace instead of RGB, skipping the RGB -> YUV conversion that was previously required. The software is also much faster. However, processing of CX-compressed analog audio still requires improvement.

Another fascinating development is the application of AI to anime. For decades, AI was a pipedream, and even today, it requires massive computing power. However, today's gaming computers are monsters: 8 to 32 cores, 16 to 64 threads, all the memory you can eat, and ultra-fast NVMe-based mass storage. Open source AI software is available. One of Orphan's encoders has been training an AI on hand-drawn anime from the 80s and 90s, and he is seeing substantial image quality improvements. On a commercial scale, AI is being used to produce better upscales of standard-definitions sources to HD.

Of course, there are severe limits; it's still early days. The underlying sources have to be free of mastering issues such as blended frames and interlacing issues. The technology can do cleanup when none is required or desired, so manual inspection of results remains a necessity. But the promise is there. Someday, it may be possible to undo the problems in even badly damaged sources.

One development remains tantalizingly out of reach: a "Domesday Duplicator" for VHS tapes. This would tap into the output from the helical scan playback head and then do all the decoding in software. It's a very complicated problem in both hardware and software, but it would get around some of the limitations in uncompressed capture, such as timing issues that require a Time Base Corrector. Hardware that would allow the tape to be "oversampled" as it is read might help even more.

Do all these improvements presage yet another round of Orphan re-releases? Thankfully, no. The initial jump in video quality from years-old Internet raw to Domesday Duplicator capture or uncompressed VHS capture was massive; just look at Meisou ou Border (VHS) or Al Caral no Isan (laserdisc) if you need convincing. Further improvements are more incremental, and they'll be applied to new releases, not old ones. Unless, of course, AI can someday undo all the frame blending in Amon Saga or Every Day Is Sunday... just kidding.


1 comment:

  1. This is certainly a very exciting time to be watching this space. Work is being executed on "vhs-decode" - https://forum.videohelp.com/threads/394168-Current-status-of-ld-decode-vhs-decode-(true-backup-of-RF-signals)

    It's also interesting seeing AI upscaling beginning to be applied to older SD media, especially at a commercial level.

    ReplyDelete