Tip: Three step tuning method, to recover bad quality videos

Video AI General
21

I use FFMPEG to resize to 720 using CPU only x264 encode at CRF 14 Superfast.

22

Yup, I always have to account for the sharpening that happens. If I am running a source through topaz that already has pretty crisp edges I usually will do light a light dehalo pass first with Staxrip or Hybrid to soften the edges up a little first.

23

Hi salva
I have only one Proteus in the Settings, I’m using Program Version V3.3.10

24

Just finished trying this on some VHS footage. I cannot say that it came out any better than QTGMC deinterlace denoise → Proteus Manual. In general, the faces looked the same, but the background got random blurry patches with Iris.
It was loads of fun to try though!

25

thank you


26

Hi
 I want to try your method. Please can you tell me what is your setting for recover original details?

27

Hello. Tried this multi pass process now several times. It gave me very different results following the initial three steps. So I used iris and Gaia. It looks imo way better than results before. Only thing is that I can’t get my encodes as sharp as I wish
 It looks either a bit to blurry or it is oversharpened and contours look a bit to strong. Maybe I’ll find the sweet spot

28

The default (which I think is at 20) for recover details. After trying some of these other methods, I’m having better luck with denoising as a first step and then working with the different models depending on how the source is reacting.

29

The thing is and a a write above, I have sometimes problem upscaling with Proteus, because I get artefacts and Iris is not my thing because it’s less sharp. When I use Proteus (relative to auto, moderate settings) without upscaling and made the upscale with other model for example “Artemis Strong Halo” I don’t have this artefcat problem and get a sharp result.

I know “Artemis strong halo” could be very difficult depends on the source, but with good denoising and sometime I do a re-denoise the Proteus Result at low value (value 5-10) before use Artemis, then often I can use the Artemis upscale.

Artemis has a effect no other models have, hard to explain but you get more shiny effect, it does a form of “wood stain” or “transparent varnish”. Some vids I get amazing results with this method, so often I try out this first.

30

This is very helpful, thank you! I am working on a VHS tape of Easter, 1985. My uncle gave it to me to digitize and I’m excited to use video AI again. I haven’t messed with it since last winter. This is the perfect project for it.
One thing I don’t fully understand is why Topaz says the video is 1,000 fps. I know it’s interlaced so I used the “deinterlace footage to upscale to HD” preset first and that can = high fps. But the framerate under the output resolution must be 1000 fps. I tried 30, then 60 but it’s jerky like the fps is actually 2 lol. If I keep it at 1000 fps it acts normal. I suppose this might not matter because Bitrate wins in the end.
And it can’t really be 1000 fps, my monitor would tear and have artifacts wouldn’t it? Its refresh rate is 120hz overclocked. Pretty far from being able to push 1000 fps.

31

I had some nastily small postage-stamp sized clips and my default SR preset setting of auto + 100 for improve detail and sharpen, and recover detail somewhere in the 20-100 range just produced distorted faces. However your tip of reducing the relative auto strength made a world of difference.

in fact using your exact values in “Recover detail” (auto+50), “Improve detail” (auto+50) and “Sharpen” (auto+70) provided amazing result using iris-v1 (a.k.a. “Low Quality”)

This just goes to show that

  1. Ones magic preset that worked so well on some sources, may be totally unsuitable for others (as in my case),
  2. That Iris is a very different model from the others, and
  3. To not give up on Iris, because some adjustments to the hyper parameters (ie. sliders) can make a massive difference. As in this case; from completely unusable result to rather good and definitely usable.

I’m still astounded how good Iris can be. Often what it achieves seems impossible when looking at the neighboring frames and observing the detail it produces just isn’t in those frames. At times, like this one, no matter how much I tried to mentally project pixels across frames, I just can’t figure out where it gets those details from. My best guess is some combination of “inpainting” using statistical probabilities of what “typically” should be in spots of a frame that it couldn’t fill with info from neighboring frames. Regardless of how, the result seems like magic.

32

I had such crazy high fps already. Often the reason is, container format contains wrong metadata. I recomment using free Muxer like “TS-Muxer” and set fps value to correct value. After that you can use for example “My MP4Box” Muxer to make a mp4 format again. All this prozesses are lossless for the video and audio stream.

33

Hello Again Everyone,

DVD → 4K

I love doing these multi step processes and have tried a lot of variations over the past 6 months. I’m now working on upgrading 1980s music videos from commercial DVDs (ripped .vob files). I can easily spend 1/2 hour tinkering on a single music video but have about 800 I want to to do so wanted to hone in on a streamlined process. I’m sharing for anyone who might be interested in getting DVD .vob files to 4K. (I know people like to deinterlace before pulling files into Topaz but I have gotten better results using Topaz to deinterlace in the 1st step).

After trying many different model variations and steps, this simple 2-step process yields incredibly good, balanced, results across multiple sources:

Step 1: Pull the .vob file into Topaz, keep Output Resolution to Original (no enlarge), set Video Type to Interlaced (under Enhancement) and use Iris MQ with Parameters set to Auto. (Keep recover detail at 20). I set the Frame rate to 2x (optional).
Step 2: Take the output from Step 1, and set Output Resolution to 4K. Video type is now Progressive and use the Gaia model.

That’s it! Hope this helps someone out there.

-Luke

34

You might get better results if in step 2, you would use Dione TV (or DV), instead of Gaia. it will be cleaner, faster and bit sharper and less blur.
I already tried your method and I found the use of the Dione AI (TV / DV) as the second step yields better results then using Gaia.
but that depends on your source video. sometimes the more blur is better as blurriness covers up scratches and over screwups of the original video.

Also there is a very good chance you might find that de-interlacing with QTGMC (using Hybrid) your “vob” file 1st and then doing step 1 & 2 with the de-interlaced file would give you better results as apposed doing it directly on the “vob” file.

35

I tried these methods on one of my more grainy shows. Dione, was way better than Gaia as the second enhancement—but both were not as good as only Artemis HQ for that show.

36

you mean Artemis HQ was the second pass or only pass / direct upscale?

37

Yes when video is interlaced, some DVDs video it’s prossesive

Since Iris, my thread is outdated! Here is the revided version of the procedure for special use, to get to high contrast/shiny results:

Step1: Use Iris
Upscale with Iris to desired endresolution, because this model does this very good. Setting “relative to auto” and set parameters lower than you normally do! Give more sharpness (when source allows), also depends on how high you upscale.

Take care on denoise, which is as always one of the most important part. Check darker scenes and denoise until most of the “fog” is away but take care most of fine structures/details are not denoised away.

Step2: Proteus (why? read note at the end)
Take the result from Step1 and use Proteus. Previews use “Auto mode” first and check result. When ok start, when not use “relative to auto” mode:

  • Need more sharpness encrease it, need more denoise, do it.

  • “Improve Detail” and “Anti-Alias” none, or at low to moderate settings (we have relative to auto mode on) because we did a first pass with Iris already.

  • Check if there are still compressed artefacts, rise “fix compression” only as much as is necessary. To do that, look at dark passages to see whether there are still compression/pixel-blocks. A little bit blocking is ok but to much you have to use “fix compression”

  • Something about “Halo” parameter. I thooth don’t touch but I was wrong. I had many cases a little bit (1 - 3) “Halo” the picture improved and to my surprise, sometimes even this gives the image more depth. This happens in cases edges are too hard, shadows are too roughly graduated, when the image is not soft enough, then it’s less spatiality for our eyes/brain. Additional “Halo” alleviates the problem of faces looking too distinctive / disfigured from a distance.

Step3 (optional)
Take the result from step2 and make a third run with “Artemis Strong Halo”. I know the model is tricky, but no other model does what this one does so well: burn in effect, It glazes as if with a varnish and that leads to more shine and also brings about a little more sharpness. Just try it out, it’s not suitable for all cases.

Note: Why both, Iris and Proteus?
Iris does a better job on upscaling than proteus, and allows high values on “improve details”. The lack of Iris is, the model was also made for low quality sources, some lack of sharpness and it’s hard to get the rest of noise away. With Proteus we ripp the rest of fog away, we use a second method for sharpening and have all the abilities to fine tune the result.

38

Thanks for the suggestions Akila. I have tried Hybrid several ways and it seemed to be good but I like the Iris MQ output better (and its also one less step!). I think my use of Gaia to enlarge is more of a personal preference.

  1. I’ve been using Dione DV like you suggested as my go-to for a 1-step enlarge result on deinterlaced material. It does really well with DVD source especially when the original source was film (rather than video tape). For retro music videos and 1980s video material (like sit-coms), I find Iris MQ works wonders to clean up and deinterlace but when also using Iris to enlarge I often ran into artifact issues. That’s what led me to doing a 2-step process.
  2. Gaia is just my personal preference for enlarging. You could definitely use Dione DV/TV too. I like Gaia because it keeps depth and realness. I even prefer Gaia (over Proteus) when upscaling movies off of Blu Ray discs to 4k because Gaia retains original qualities with subtle improvements, but without looking filtered. Using Dione DV/TV or Proteus to enlarge the output indeed looks stunning but not quite as “real” to me as Gaia.

What’s great about Topaz is having all these options and abilities to fine tune.

39

Thanks Mayday! You’re multi-step approaches have helped me approach lots of differing sources of material that I had initially given up on. I’ll definitely try this out on my progressive videos. The new Iris MQ is really great, too.

40

Artemis HQ as the only pass. It works really well on that specific show, but not most of the others I have tried it on.

@Mayday, I like where you’ve gone with those latest steps. I’m going to pass on trying them though. The rumors of Proteus 4 give me hope that it may be a really good tool.