Teranex AV - not happy with results

So I've had my Teranex AV since a month and have now done a lot of work with it.
The main purpose was to convert broadcast MXF 50i to 50p and upres old archival SD formats to HD in an easy way: press play and record the finished output.

My main problems:
1) old SD footage with lots of noise - the denoising quality is bad - it looks like a simple temporal smoother with all its disadvantages (like pixel-ghosting and pixel-smearing). Compared to what you can get from NeatVideo it is a complete joke...
The denoising algorithm seems completely unaware of the pixelmotion vectors or any rainbow cross-talk of old tape footage.

2) Deinterlacing results are disappointing to say at least. Some of the old footage had been originated from 25fps BW film and been upscaled to 50i HD with some thrown off 2:2 cadence. The Teranex often completely failed to recognize any traces of interlacing. Even throwing away one field in VirtualDub2, downscale and upscale in the edit suite resulted in better quality than the Teranex directly.
Deinterlacing of modern HD 50i also often failed to remove the fields or resulted in jaggies.
I get superior quality by using freeware software tools like AviSynth script QTGMC in StaxRip.

3) The Media Express software is also pretty useless. It doesn't play MXFs, it doesn't play ProRes or DNxHD MOVs or any other format you would work with in a professional setup. In the end I used Premiere Pros source monitor to play out the files to a UltraStudio Mini HD, from there to the Teranex AV and record the output on a Shogun Inferno.

Some examples for deinterlacing problems:

Source_50i camera starts to tilt down the building

Source_50i_0000300.jpg (687.86 KiB) Viewed 3858 times

What the Teranex makes out of it

TeranexAV_0000600.jpg (541.2 KiB) Viewed 3858 times

What the QTGMC software deinterlacer makes out of it

QTGMC_0000600.jpg (540.99 KiB) Viewed 3858 times

At the end of the tilt:

QTGMC_0000800.jpg (644.12 KiB) Viewed 3812 times

TeranexAV_0000800.jpg (680.05 KiB) Viewed 3812 times

Source_50i_0000400.jpg (780.64 KiB) Viewed 3812 times

Then in this sample look closely at the details of the jackets and the key and the pattern of the tie which was diagonal from left down to right:

Source_50i_0000630.jpg (655.31 KiB) Viewed 3809 times

TeranexAV_0001260.jpg (432.56 KiB) Viewed 3809 times

QTGMC_0001260.jpg (406.69 KiB) Viewed 3809 times

And finally a quite static interview - look closely at the hair, the shoulder lines, the station logo on the windshield, the face details:

Source_50i_0000846.jpg (562.15 KiB) Viewed 3809 times

TeranexAV_0001690.jpg (483.84 KiB) Viewed 3809 times

QTGMC_0001689.jpg (472.48 KiB) Viewed 3809 times

No surprise. I have not seen any box outperforming QTGMC even with its default settings/presets. Once you add its all possibilities then it's simply masterpiece when it comes to deinterlacing.
Teranex is more on the YadifMod level than QTGMC.
Face details are actually to sharp (QTGMC by default uses small sharpening). I would actually lower it in this case. Where Teranex fails a lot is diagonals. As you said- sometimes it simply fails to perform deinterlacing and leaves combing. Bottom line- don't believe that something marketed as pro has to be great
Snell's software version of Alchemist is definitely better than Teranex, but occasionally also fails to remove combing. This happens only when moving parts cover tiny areas of the whole frame, not like in your examples.

Yeah, I knew I would get better results by hand tuning software filters, but with 50 clips lasting several hours I couldn't afford this time consuming task anymore.

After inspecting more clips I think it seems the Teranex algorithm completely fails to deinterlace when there is a straight up- or downward tilt of the camera. This should really get addressed by BMD.

I've put almost 1000 hours through QTGMC recently
Use faster presets- they are still very good.

The problem is that the archival footage is just a mess, sometimes SD, sometimes SD in HD, some have the fields set wrong (and this is from a big public broadcaster). Some would need dehalo. Some have the aspect ratio wrong.
Footage originates from BW film, color film, Umatic, Digibeta, old JVC whateverformat.

Then I would like to keep TC and the 8 audio channels intact - I haven't found an easy method with StaxRip to do that and while years ago I used to write AviSynth code, it is frustratingly time consuming to dive into that again. All while nobody is willing to pay me for that time and effort.

When I update our post hardware to AMD Threadripper I might use QTGMC more often again - ATM it blocks too much of the available CPU resources.

I have very custom made setup for this, which also uses power of Vantage and Python. It's fairly complicated system in some way and so easy in other. All works very well and been rock solid.

You should set the source type to "video" and stay within the same family of framerates to achieve the best deinterlacing results that are possible with the Teranex. 50i to 25p or 50i to 50p will be fine (50p will keep more temporal information in this case).

Deinterlacing always stays within same fps, otherwise it's fps conversion not deinterlacing.
Video vs. data levels will just give to bright/dark video. Won't really affect quality of deinterlacing itself. If it does then it's massive bug.

The Teranex deinterlaces AND converts the framerate at the same time if you set it to.

But it keeps only the maximum vertical resolution if you keep it in the same family of framerates (25/50, 29.97/59.94, 30/60). As soon as you use the framerate conversion, the vertical resolution drops noticeably.

Source type Video, Film or Auto does not set the video levels, it configures the cadence detection.

The Video setting is much more reliable with difficult material.

Film or Auto should only be used with 29.97 to 23.976 conversions from film, which has been telecined to 29.97 using 3:2 a pulldown, this will not degrade the vertical resolution as long as the 3:2 cadence is relatively clean and there are no exotic cadences used.

Different thing then, but I'm sure Robert was using same fps.
Video will most likely blindly assume source is interlaced. 2 others will be adaptive and try to analyse source before making decision.
I know Teranax advert page- converts to hundreds of formats, I would prefer if it could do 10, but properly

Markus, I did tests with the settings during my preparation of the setup, but I found that the setting of the cadence detection had no influence on the outcome as long it was set to auto or video.
I will do some further tests, maybe I just had bad luck.

I am using the Teranex 3D. The conversions should be the same on the AV, but I cannot tell for sure. I used it to convert lots of mixed materials (all-interlaced video sequences together with 3:2 film based material) from 29.97i to 29.97p and 59.94p.

It was set to video as this gave the best results (auto tries to switch automatically between video and film guided cadence detection, which often fails on difficult material).