Solution for Weird Faces, Upscaling from LQ to HQ upscale without weird faces

Hello all,
I have found the solution for weird faces when upscaling. I had been looking this for a while, finally found it. Here is the sequence I took to get great quality without distorting faces:

A. Deinterlace video with Hybrid ( a free Windows program, but you can use on Mac with use of virtual machine like Parallel desktop etc). or you can use TVAI Dione DV3 to Deinterlace (Hybrid better > Dione DV3).

Hybrid settings:

  1. use video = FFV, container = Avi , audio = passthrough for output:
  2. Crop/resize: use x2 scale, resize method = 10-bicubic spline, check the auto adjust, use width x 2 from original source:

  1. use Deinterlace custom setting, with QTGMC , here are the settings:

  2. use Filtering → Vapoursynth → Resize : use resizer NNEDI3, use neurons=256, neighbourhood=32x4.

Bring unmodified video ( do not use Handbrake at this time), and use it in step B.
B. Use Progressive, Iris V1 with recover details 100, and all settings O and Denoise to 100. you can upscale to x 4 or 4K or 8K.

I had issues with faces, but finally do not see any weird faces with above settings.
Enjoy!!

1 Like

The faces look weird because you forgot to choose Squre Pixels when upscaling dvd/mpeg2 video in Topaz (models were trained on Square Pixels videos).
Try upscaling your deinterlaced/detelecined dvd in Squre Pixels mode to 720p with Proteus v4 Auto with 3% Original Details and then run the result (ffv1 lossless) through 1x(no upscale) Artemis v10 High Quality with 100% Original Details for extra sharpness. To my eyes this produces the least amount of distortion on a full hd screen.
Then you can encode ffv1 to x264 profile High 4.1 CRF 12, very slow, tuning None if you wish to save space.
You have to accept that it is impossible to upscale a dvd to a native 1080p video, there is simply not enough details in the source. The higher you go, the higher is the distortion/visible upscaling errors, that’s why I settled on 720p for dvds.

So I tried to use the Hybrid and now it is rendering for more than 3 hours now and I don’t even understand what is going on why the process is going over 100% now

It is also extremely tedious to use this program and the UI is horrible, not user friendly at all.

I choose this container because my source file was mp4.

I don’t think it could serve as a servicable solution.

I have seen this too on my videos as well, it takes longer as the settings of speed is very low, but ultimately done with good quality lossless video. unfortunately , it takes longer time, may be they have issues with the UI.

Hopkins, your Hybrid screenshot indicates that you chose slow lossy x265 codec instead of lossless and fast FFV1.

1 Like

Thank you for the suggestion. I will try the square pixels as well. My DVD with Topaz 8K shows amazing result with the settings mentioned above . I never had that experience with topaz and even 1080p was giving distortion but I am really happy with upscaling to 8K after Hybrid settings and just using the denoise only option with Iris V1.
I might experiment more, with sharpening after downscaling with handbrake.

I’ll try your settings but I think it’s a lost cause. Iris makes you lose too much detail. In QTGMC on your screen you activate “source matching: x2 refined” but it is linked to “match enhance” and you left it at 0 so this setting doesn’t change anything. So basically you use these settings with hybrid exported in ffv1 and you import this sequence into VAI to use Iris? And in my SD videos (720 x 576), putting it in square pixel or not doesn’t make any difference in quality

2 Likes

yes, exported from Hybrid with x2 upscale FFV1 file, import that to TVAI with Iris1 just deniose to 100 and recover detail 100. (I checked all with Denoise 100 and recover details with other models but I found Iris 1 the best out of all). My DVD videos look great with above settings. Before same faces were distorted even with upscaling to 1080p with other TVAI settings. I still have to do work on my 8K videos to downscale and run other settings to see how the final product.

Hybrid was used for Deinterlace only (with upscale x2) without changing anything else (minimum original video change). TVAI is just denoise only.

ok, i think that even if it works well for your dvds, for me it’s another quality, my sd videos are digitized 8mm tapes, so there is no compression but a lot of noise with artifacts, and in a case like mine iris does less well than proteus. On the other hand, using “very slow” with qtgmc makes you lose a lot of details. even from “fast” the loss is visible.

1 Like

why not already convert your video to Square Pixel while you at Hybrid doing the resize (screenshot 2.)?
that is what I do, I convert to square pixel in Hybrid as part of the deinterlacing

Thank you for the tip. I was trying to retain as much original details without messing many options. same video anything can be done later on. Not sure, how much difference square pixel vs original pixel would make though.

  1. you do resize from 720x576 to 736x576 to bring 5:4 to 4:3
  2. you do again 720 to to 736, but we have already 736?

But you can do this also by saying output “square pixels” in the deinterlace register. Or do you 2x upscale with Hybrid? But Screenshots does not show that.

I know “spline64” is optimal doing a scale without halo. But in general I’m not sure if it really generates better results when you scale in Hybrid compared to import deinterlaced orig 720x576 into TVAI and set there “square pixels” which gives you also native 4:3

QTGMC deinterlacing in Hybrid: When you set the “BOB” flag, this does double framerate. I did it and compared 50fps and 25fps deinterlaced results.

I noted the additional BOB generated images does often not looks good, so then I decided not doubling framerate while deinterlacing. I recommend take a frame by frame look on the result and then decide BOB makes sense or not. In my case I did not, later I would decide using Topaz AION model for frame interpolation or not.

Unfortunately screen shots do not show actual process being done, as they were taken after the process was done. I did scale x2 (720 x 576 to 1440 x 1152) and also did original resolution as well. I have found Scale x2 imported to TVAI and use Iris V1 with mentioned setting had better result. With recommended settings the process in Hybrid took long time though.

You “found the solution for weird faces when upscaling.”? Really? That’s a wee presumptuous, don’t ya think?

Also, “Deinterlace video with Hybrid”? Why do you assume everyone has to deinterlace?

I appreciate you’re trying to be helpful, but finding ‘the solution for weird faces when upscaling’ is, frankly, as good as totally out of our hands. Topaz can improve the models, and we can wait for that to happen, that’s it. Of course, we can tinker a bit with sliders, to maybe offset weird effects from occurring, on an individual basis, but there’s no panacea, aka, no overall ‘solution for weird faces when upscaling,’ other than Topaz continuing to do their thang.

1 Like

fixed screen shots.

I agree with you , technology has to evolve and it is case by case. I used to get very weird faces, and struggled for many months to get better results and with above settings at least my videos got much better, reasonable and thought it might help others to try similar settings. Of course, whoever has interlaced videos, like mine or home videos or older videos are interlaced, and only those needs to be deinterlaced.

And sorry if I sounded a bit harsh there, as I really wish you actually had found ‘the solution for weird faces when upscaling.’ :slight_smile: But every bit helps, I suppose. :ok_hand:

This is what I do (albeit it not in Hybrid, but just directly via my own VapourSynth pre-processing script) . QTGMC is still the best, IMHO, but yeah, doubling the frame-rate, especially for 4k end-result movies, can become quite taxing on your media player.

And yeah, always use lossless (or ‘near lossless’, like 422 HQ .mov) formats for your intermediate files.

qtgmc renders cleaner footage with less artifacts compared to others, but I noticed that it still renders less detail compared to yadif or muksun with virtualdub