I found a flow that I think managed to yield very good quality on my SD footage (Video8 PAL 576i)
I want you to try and let me know if you find this flow works for your stuff as well.
deinterlace with Hybrid at Slower preset (or use your favorite Hybrid settings) + Converting to Square Pixel - Export as Lossless compression (FFV1).
If your video is already at PAR: 1x1 (Square Pixel), then no need to convert it, only deinterlace.
load the deinterlaced / converted FFV1 videos into TVAI and run it with Nyx AI (same resolution - 720x540) - Relative Auto and change only the Deblur to +60, the rest I left at 0 - Export as FFV1.
a. Load the FFV1 Nyx treated video(s) into TVAI, Upscale x2 using Iris v2 (Auto) + Motion deblur (Themis v2).
b. Alternative option (Step 3), is to Use Gaia HQ instead of Iris v2 + Themis v2.
it provided me shaper output and very detailed output, but did not cleanup some artifacts i had on my original videos.
Please try Step 3 both with Iris v2 + Themis (a.) and another round of testing using only Gaia HQ (b.).
Please try it out and let me know how it works out for you…
I’m going to try your flow, I have the same SD videos as you I think. I tried to do like you yesterday, combining nyx with artemis hight and gaia before the upscale with irisv2, but I haven’t tried themis yet, in the past it gave me artifacts. and then the difference is that I abandoned the square pixel (I set the source to 768x540) with hybrid some time ago because I found that the quality with vai was the same, but took longer to process . (with the resizing of your video, you use lancsoz?) and also, you deinterlace using klnmeanscl? keepgrain or ezdenoise? approximately what is your amount of noise?
You are correct, it used to give me artifacts as well, but for some reason I don’t see it happening that much with Iris v2. some times I noticed it cleanup the video better then just using Iris v2 alone. very strange.
Is there a better recommended resize method to use? , it is more of a downscale then upscale. 720x576 → 720x540.
I just use the OOB Slower preset. it reduces some noise, yes, but i found it not to be to aggressive.
agree for themis, strange indeed but so much the better! for resizing, yes even in sd we see the difference between lancsoz, bicubic, bilinear, spline, area, sinc etc… you should definitely not use a default preset in my opinion because already, with the slow preset, “final temporal smoothing” is at 1, even if the grain is smoother, we lose detail in the movements, so I always leave it at zero for my part because I think that “vai” with all the models denoises better than qtgmc with its denoiser, and then even if you want to keep the preset in slow, I advise you to put “custom” to leave everything in slow but to define the denoiser “klnmeanscl” instead of “auto” which is “dfftest” KNLmeanscl performs better and loses much less detail.
So which Resize method you recommend using out of the list there?
What are the settings you use for QTGMS?
There is one I tried before was from @ForSerious , But when I ran upscale tests between the “Slower” Preset and the one from @ForSerious , I noticed that the quality of the Slower was slightly cleaner.
I’ll re run it again and see if I feel any differences. please post your settings as well and I’ll try and compare them again on my videos.
It’s normal that it was cleaner since your temporal smoothing is at 1 and therefore less noise in the movements and much easier to have a clean video for upscaling. the disadvantage, the more we increase, the less precise details we have on the movements, in fact it goes like this (I set the keepgrain to 10 to see better by raising the noise because at 0 or with ezdenoise, we distinguish not easily!)
obviously we have much less noise but too much loss of precise details for my taste… the objects or the vegetation, the bricks or the cracks on the walls suddenly seem unreal
This is not the best example because the movement is not very important in this regard and by improving the details and the sharpness, once the video has been retouched, we notice better!
I’ll try those settings along side @ForSerious settings and run a compare on my final workflow and see if I can spot the difference.
in the mean time test my workflow (Step 2 & 3) and tell me what your take on it.
I tested your settings and also @ForSeriousSettings compared to each other and compared to OOB “Slower” Preset.
and the differences is so minor (after running the 3 step TVAI process) that it took me long time to spot the difference in detail gain.
I used the Vide-Compare tool for it as i analyzed frame by frame.
In either way the “Slower” Preset provided me the cleanest final outcome after running it through Topaz (my 3 steps process) and the level of details lose compared to your QTGMC (and @ForSerious ), was so minimal to hardly noticeable, but the cleanliness of the “Slower” was more noticeable then detail loss.
My conclusion is the added value of the cleanliness of the videos outperformed the level of detail gain with with your (and @ForSerious ) QTGMC settings.
So I will stick with the “Slower” OOB Preset/template that QTGMC offers.
I am confident that my settings can be improved. Good work finding something better for you.
I’m also super interested in how this flow works on my VHS recordings. I thought I had a small clip already prepared in Hybrid, but it didn’t turn out good at all so maybe it was a different thing I was testing. I’ll see if I get time later today to do it right.
so you prefer cleanliness to the amount of detail but with good, perfect settings (which I haven’t quite found), you can have cleanliness with all the details. I tried themis on artemis hight, gaia, iris etc, each time it gives me weird artifacts on certain scenes and I can’t remove them even if the video has no noise.
it might work as mines are Video8 which is not so far out from VHS
no, I prefer details, but the level of detail was so minor to hardly noticeable when I used your QTGMC settings compares to using the Slower Preset. but the level of cleanliness was much predominant and noticeable. So cleanliness supersede in my case as hardly any details were lost vs. level of cleanliness…
Another thing you can try (i tried it after posting), is in Step 2 where you do the Nyx, do Nyx + Themis and in Step 3, Iris v2 - just keep on auto (without Themis). this could help if Themis introduce artifacts, the Iris v2 pass will clear those out, but you would still get sharp video. not as sharp as the original flow i posted, but sharp enough.
here is a summary of my finding i captured in Notepad
Flow 1: DI Hybrid (Slower preset) → Nyx Deblur 60% output same Resolution → UPSCL x2 Iris v2 Auto + Themis v2 Or UPSCL x2 with Gaia HQ without Themis
Flow 2: DI Hybrid (Slower preset) → Nyx + Themis v2 output same Resolution → UPSCL x2 Iris v2 Auto Or UPSCL x2 with Gaia HQ - NOTE: UPSCL in this flow with Iris v2 little bit cleaner then Flow 1, Use only if Flow 1 has artefacts. UPSCL x2 with Gaia HQ much Sharper then Gaia UPSCL x2 in Flow 1, same quality give or take, and more details preserved vs. Iris v2. but could not fix faces in poor situation that Iris v2 does.
I agree about the cleanliness, but I still think that setting final temporal smoothing to 1 instead of 0 like your slow preset does makes the video less natural (as if it adds motion blur then details less visible), worse still, the fact of not using klnmeanscl and using dfttest (which is the worst between the 3 because dfttest removes too much detail at high frequency), it loses too much detail in the end. but if you prefer cleanliness, without taking too much trouble keep the preset slow.
For my part, my videos are very clean without noise, but I have artifacts that did not originally exist because of iris v2, that’s the only problem I have. I tried to do a pass with gaia hq in addition to iris v2, the video is even more beautiful by smoothing the jagged edges and removing the last residual noise but it amplifies the artifacts caused by irisv2, which is a shame !
I’ll tweak nyx with themis then and do more tests
There you go, too much loss of detail, because of nyx which removes all the details, particularly on the hair here, and themis which deblurs well but smoothes the image, which gives a totally plastic appearance.
I also tried nyx 60% + irisv2 + gaia upscale x2 (horrible):
for resizing, yes even in sd we see the difference between lancsoz, bicubic, bilinear, spline, area, sinc etc… you should definitely not use a default preset in my opinion because already, with the slow preset, “final temporal smoothing” is at 1, even if the grain is smoother, we lose detail in the movements, so I always leave it at zero for my part because I think that “vai” with all the models denoises better than qtgmc with its denoiser, and then even if you want to keep the preset in slow, I advise you to put “custom” to leave everything in slow but to define the denoiser “klnmeanscl” instead of “auto” which is “dfftest” KNLmeanscl performs better and loses much less detail.
I have improved my workflow quite a bit like that
