Piracy: ꜱᴀɪʟ ᴛʜᴇ ʜɪɢʜ ꜱᴇᴀꜱ @lemmy.dbzer0.com Lemmchen @feddit.de 1 yr. ago

What is the current state of the art of upscaling DVD-only releases?

I know this technically doesn't really qualify as !piracy@lemmy.dbzer0.com, but I figured if anybody on Lemmy would know, it's you guys.

I have some old DVDs that never got a proper Blu-ray release. With all the fuzz about and the recent advancements in AI, what is the definitive best way to upscale to 1080p in 2023?

20 comments

I do a lot of AI upscaling, specifically 1080p to 2160p with Topaz VEAI, and its quite good. That said, it isn't magic. It requires a really good source before hand (avoid anything with film grain or has been compressed) and you do not want to make a giant leap. The key idea is the more you tell the AI to fill in the gaps, the more you will get AI shenanigans. If you have a good clean BR rip, then the AI can bring it up to 4k quite easily.

what is the definitive best way to upscale to 1080p in 2023?

1080p to 4k is quite easy, but getting something to 1080p is massive mountain to overcome. Like I said above, you want to leave as little to the AI as possible. Bringing something sub-HD to HD means entrusting the AI with a lot since there isn't much detail to begin with. My best efforts have been 720p-1080p upscales, but I have had some good results with upscaling the 1990s Sabrina live action series and off the top of my head that is 480p. If you look really hard on those upscales, you can see some strangeness like how words in the back ground are just gibberish or maybe a hand doesn't move exactly right, but it is a vast upgrade over the terrible DvD release with its low resolution and interlacing.

If you are going to upscale, I recommend Topaz VEAI. Its really easy to learn (I went from having no video editing experience to being able to do some competent upscales in three months), its a lifetime license with a year of upgrades, and frankly its the best on the market. The only drawback is like any AI, you need a hefty rig. I am rocking a 3080ti, i9 processor, and 128gb of RAM. Its also very time consuming. My hefty rig can do the upscaling part in about 45 minutes, but even taking the mild tweaking in DaVinci out of the equation, encoding that same episode is going to take a good 11 hours unless you want a 200gb episode on your Hard Drive. (Please do not trust Topaz to encode for you, its a good upscaler but its better to let Handbrake do your encodes). Throw in the deinterlacing I do before I even run it through Topaz, and it can take me 24 hours of work spread out over a week to get an episode done.

If you have any specific questions, I will be happy to answer them.
- How much RAM do you need for Topaz? I have access to a 2080 ti 11GB and a 6900 XT, but both systems only have 16GB of RAM.
  
  Are there any specific profiles for Topaz you can share? (I assume it supports something like that?)
  
  Is there something equivalent to Topaz on Linux?
  
  How do I get a DVDRip to look clean without any further compression? By default MakeMKV results in files with these annoying lines, which only disappear after passing it through Handbrake (deinterlace/decomb).
  
  Will I have better results only upscaling to 720p or does the output resolution not matter in terms of AI artifacts?
  
  How much RAM do you need for Topaz? I have access to a 2080 ti 11GB and a 6900 XT, but both systems only have 16GB of RAM.
  
  Before I built this machine, I had 32gb of RAM. You should be fine with 16gb as it isn't that RAM heavy. Just close out anything not Topaz. Topaz is going to 100% your machine (like it does mine) anyways so you do not need anything else up. On that note, I recommend upscaling over night. A 1080p-4k upscale can easily take my machine 7-11 hours to do. You shouldn't have that long since you are only trying to bring something to 1080p, but still be aware your computer may be occupied for a couple of hours.
  
  Are there any specific profiles for Topaz you can share? (I assume it supports something like that?)
  
  Yes there are profiles, but they are better for keeping things persistent between series (like I have a Sabrina profile just for Sabrina) rather than a one size fit all. I do have one I found to upscale BRs to 4k with only slight adjustments. I will post in a follow up comment once I get off work. In general though you want to make slight adjustments, render a few seconds of video, review it, then make more adjustments. When you like it, apply that to different scenes in the same movie (I recommend testing dark/night scenes for sure since those a tricky for AI) to make sure it works, then do the actual upscale.
  
  I will also throw in my Sabrina preset since I managed to use it to upscale some other DvDs with minor tweaking as well. It should provide a good starting place for your DvDs.
  
  Is there something equivalent to Topaz on Linux?
  
  I do not know Linux, sorry :(
  
  The amount of commercial AI upscaling software is rather limited (at least when I started searching this May) and focused towards a Windows and maybe Mac audience. That being said, I know there are AI upscaling software on Github so I would not be surprised if someone ported something like Waifu2x (an open source AI trained to upscale anime) to Linux. As to their quality, I do not know. All I can vouch for is that of the commercial software, Topaz was the best when I bought it in June.
  
  How do I get a DVDRip to look clean without any further compression? By default MakeMKV results in files with these annoying lines, which only disappear after passing it through Handbrake (deinterlace/decomb).
  
  I do have a program to deinterlace (remove the annoying lines) which I will also provide you the name of when I give you my profiles. Topaz can also do decent deinterlacing, but my program is better (and also free).
  
  Will I have better results only upscaling to 720p or does the output resolution not matter in terms of AI artifacts?
  
  It is going to depend on how clear your source is. If your input has a lot of compression artifacts, then that means more guesswork for the AI, and more chances for Topaz to guess wrong. Upscaling to 720p is going to be a lot more successful in general since its borderline HD, but if you have a good source and don't mind the occasional weirdness really only noticeable if you know what to look for, then 1080p is doable for even something like 360p.
  
  One thing I like to do (to the anger of many film purists) is add my own film grain in post. There is something about film grain that helps our brains not notice small errors in the upscaling process. Holy Grain is a very good film grain in 1080 and 4k. Sadly I never found a torrent of it so I ended up spending the cash to buy it myself.
I upscaled some old porn movies using this, but my 2080ti started flickering afterwards. Probably better to use the free Google resources: https://github.com/k4yt3x/video2x
- Hehe, but the old myth that graphics cards degrade if you use them is a myth. I think Linus Tech Tips did a video on this and an older one. Sad that your GPU is flickering now, but probably unrelated and had happened either way at some point.
  
  Yeah, it was probably a coincidence you're right. So far I've not been able to justify the expense of something equivalent to the 2080ti. I just live with the flicker whenever I work on my Linux desktop. Which isn't often, too much work and not enough brain.
- Was it worth it?
I do this quite frequently for a certain youtube channel, you'll want MakeMKV, Hybrid, and Topaz. If you really want to go all out and make a huge project out of it you can upscale each scene individually and stitch them back together in Premier, ain't nobody got time for that though.

Rip the disc with MakeMKV.

Then open and deinterlace the file with hybrid, I use the QTGMC filter, if you're doing film content you should use ivtc. Do any other filtering or denoising in Hybrid, the video really needs to be clean before upscaling. Then export it in a lossless file format with passthrough audio.

Load that file into Topaz and upscale it with the Proteus AI model, I use the manual setting (40,20,10,1,3,-2) with 25% original detail recovery and the grain filter on a low setting to help restore what the filtering removed. Your settings will need to be adjusted for each project. Export again in the highest quality that you can, then load that file into your favorite encoder and compress it down to whatever size/bitrate you want. Topaz's encoder doesn't have any configuration options so I use Hybrid to do the final encode.
- deinterlace
  
  I don't think I've ever seen a DVD with interlaced video. Is that really a thing?
  
  They are all interlaced, you just won't ever see it unless you have a really really old DVD player that isn't progressive scan or rip the files directly.
  
  The Futurama box sets I have are interlaced and the rips look absolutely terrible if I don't remember to set the playback options right when I watch an episode.
- Hybrid
  QTGMC
  
  Is there a way to simply use ffmpeg for this step (if it has been built with --enable-avisynth)? It seems like all Hybrid does is use a simple(?) AVISynth script.
  Can you export the AVISynth script you use from Hybrid and post it here?
  
  A quick search also brought up https://github.com/hclivess/videer, which seems like a simpler deinterlace tool that also uses QTGMC.
Before you upscale it is crucial to get the best rip possible. To my knowledge the best ripping and deinterlacing process is called QTGMC.

After that, experiment with AI upscalers. But the best Upscalers will only get so far when the initial rip isnt as best as possible.
Well, I just ripped a few of my old DVDs with handbrake, and I upscaled by setting the resolution to 720p and ensuring the "upscale" cjljeckbox is ticked.

Tovmy eye, it looks pretty good on my 55" 1080p TV, and I noticed in my voddeo player it scaled to my display size appropriately. If I imply ripped it, the video would have a hugee black box around it ony high red display.

I'd say using handbrake is fairly effective once you get it off the disk, for which I used a rip to mkv app in Linux.

You've viewed 20 comments.