Found a work flow to provide better up-scale Quality

Topaz Video AI General
1

Hi All, I worked hours yesterday to profection an Upscale quality work flow and I think I found it.
I am posting this and methods I used in order for you guys to test it and validate or dispel my finding.
that way we will know rather it’s a pattern or just a coincident (since I tested it only on 2 clips).

I need you to test the same on your SD videos and tell me if you guys also see a quality enhancement with those steps..
you can use Video-Compare to compare results.

Use Case: TVAI v3.1.9.
Upscale SD 576i to 1440x1080 (results were similar to when I upscaled to 1920x1080 on all tested cases detailed below)

the following steps were performed on the original Clip before heading off to TVAI.

  1. I deinterlaced the clip(s) in “Hybrid” QTGMC + Bob with Preset: Slower, exporting it as FFV1 lossless
  2. Since my clips were at PAR: 16x15, I also converted the output PAR to 1x1 (which resulted in 720x540 output resolution). You might not need this as you might already be at PAR: 1x1
    image

image

image

As of now I’ll be referring the output clip of this step “di-clip” and original interlaced clip as “org-clip”

then i took the di-clip and fed it to TVAI and enhancing it with Dione: DV (forcing TVAI to think it is interlaced video) , but leaving it at the same resolution (720x540).
I noticed that the Dione presets are not so aggressive like Proteus and Artemis. So, I wanted something less aggressive.
Since Interlace option in TVAI, it want to double my framerate (which is fine when you actual need to de-interlace. But I am already on 50 fps progressive, because of my de-interlace I did in Hybrid), I had to force the frame rate back to 50 FPS (from 100fps TVAI wanted to frame it to), which in return switched on the frame interpellation. That was fine, no problem with that, since we don’t really need to de-interlace and we just using it to be able and benefit the Dione preset.

image
image
image

Final step I upscaled the Dione: DV output to 1440x1080 with Proteus (auto), and the results were really really good. Sharp, clear, no distortion , hardly any details lost… I found if I do the final Proteus step (as mentioned above) with “Relative Auto” and bump Deblur between 50-65, I get really sharp feed. the rest i left at “0”, i.e. “Auto” on the rest.

For me i think I found the perfect workflow that provides very good quality for SD -to->HD upscale.

I need you to test the same (with your SD videos of course) and tell me if you guys also see a quality enhancement with those steps, instead of how you used to do it up until now.

I use Video-Compare - Video compare tool to compare results.

this is the full testing i did if you want to read the full report:

P.S. i noticed that if I changed to square pixel in TVAI, the quality of the output was degraded compared to original pixel settings.
No idea why, but I noticed it many times before, I think TVAI can’t do both square and upscale and/or enhance at the same step/time that good. That is why I use Hybrid for pixel converting (if needed) and not TVAI.
I noticed this also with other Filters in TVAI, if you upscale and add another filter to be processed at the same time, e.g. upscale + stabilization, upscale + slow-motion, it doesn’t do a good job compared if you do those as 2 separated steps. that is probably true also for upscale + square pixel conversion. and i believe that is why i am seeing quality degradation if I choose Square Pixel.
I leave TVAI to always focus on enhancing/upscaling without any additional steps in the process that TVAI has to worry about…
If I need stabilization or Slow Motion, I’ll do those as a separate task in TVAI, not together with upscale.

6 Likes
2

I can test it however I don’t have any SD 576i sources only 480i and the output Frame rate in Hybrid will be 60Fps. Hope it works even so.
I will report back as soon as possible
Thanks for sharing your method :+1:

3

yes, that is more then enough. the FPS is does not effect the results as I also tried without Bobbing (i.e. kept it on 25fps) and the 480i is practically already square pixel.

1 Like
4

Hi Sir! how can we reproduce and compair if not a short test file is provided?
With losslesscut you can give us a snippet.
And no link to that additional program needed?
Cheers!

5

@ArtndFun That is the point. I deliberately didn’t provide my videos as i want you to test it on your videos to see if the method is indeed a trend for all types of SD videos or it’s only limited to my videos. I provided you the steps taken that i ran the compare with (full compare is in the DOCX file, but focus on the steps in the post). you will test it on your videos and see if the results matches also on your videos, i.e. you get better quality outcome.
if it does, then we have a global method that will most likely work for all.

The test is not to see if your machine does processing different to my machine, but rather if the method works for few types of video as a global method, focusing on SD --to-> HD.

6

To be honest i am actually working on an interlaced clip, i will give it a try : )

1 Like
7

Do i need both a 1gb program plus avi synt?
Can you please try on my 6sec clip before install all that?

Thanks! It’s a little demanding one
Original Send large files | Transfer big files free online | Sprend
My Topaztest Send large files | Transfer big files free online | Sprend

8

what do you mean 1gb program?
you can test it on your sample files. I also did those on short clips, 20 sec each, etc.

9

i tested it on your video. and it came out really bad…
maybe because there is high movement in your video (they bouncing their heads rapidly).
need more testing on other movies, maybe slow movement movies??

10

Hi! I tested your method :slight_smile: so with my sd videos, it does not work, there is still grain, but it does not move, it is static and not uniform either. I think I’ll have to review my settings on hybrid then. on the other hand I have been working for a few days on the clip like a virgin of madona, and with your method, it works quite well, only in passages there are traces of horizontal lines a little, they are a little less visible when I I add 1 in grain.
and I have the watermark with the latest versions, so I’m doing it with 2.6.4. it also works even if it sends 60 fps instead of 30, I will delete every other frame after I think. overall on the clip, it does not lose detail and it blurs better than artemis medium. there are some accentuated artefacts on the other hand, but overall, your method works well.

11

I have noticed that my method works better if the original clip quality is decent (e.g. Video8 , handy cam, etc.), but I think it less works well if the quality is poor, some old VHS takes the quality is really poor.

1 Like
12

yes, I noticed that an hour ago too, the clip I’m working on (640x780), when I produce it directly with proteus in relative to auto (I had never used this mode but it works wonders on the clip) at 1280x720, the image is beautiful but if before I do a pass with gaia hq which improves the details a little or gaia cg which smooths out the imperfections a little, the pass with proteus becomes less good. in fact, oddly enough, the video enhancement works well if there are some artifacts. I haven’t used relative to auto yet with my sd videos but I don’t think I’ll use it because even if the clip image is much sharper and there are no more artifacts, some details are smoothed out. and I don’t want to lose any with my family videos. :slight_smile: I publish an image of the clip that I modify:

madona png
madona png1920×689 44.5 KB

the image is much sharper, no more artefacts but it still lacks some details at times, and I have the impression that when I play with the sliders, there’s not much that changes.

1 Like
13

Relative to Auto is just a mode that allows you to manually control one or more of the elements manually and leaving the rest at Auto, as apposed to if you pick “manual”, which you have to adjust all the elements manually.
That makes life easier. if you trust TVAI judgment on most elements but you want to tweak one (or more elements) to your liking you can user it.
For example, I noticed on my videos “Auto” settings in Proteus, TVAI does a very good job cleaning up the mess in Auto mode, but the picture is not sharp, it’s a bit blurry, TVAI I guess is not so good at blur/sharp judgment in Proteus Auto mode… so what I do is I select relative Auto, leave everything on “0” (meaning those are left at “Auto”) and only bump up the Deblur.
That way I enjoy both worlds. TVAI auto adjusting my video cleaning up the mess based on it’s judgment and I get sharpen image/video as I override the deblur (only) settings.

I would recommend you try doing the same video you just did (in your picture), but this time select relative auto , leave everything at “0”, apart from deblur, bump it to “50”, check the results, I am almost certain you will be very pleased with the outcome. you can do a quick 5 sec preview if you don’t want to encode the entire video. let me know how it went…

image

14

indeed, then in the original resolution, the difference is not too visible but when I upscale to x2 (1280 x 960), with the deblur at +50, it’s less blurry without adding any artifacts then I like it :wink: I imagine that the difference must be even more visible in fhd or in 4k… I tried proteus auto in 4k upscale but even if the image is beautiful, it lacks more and more details as the resolution increases.