Whis is Topaz Video 3.x quality so BAD compared to 2.3? 🤦🏻

Check this comparision:

Original is a 720x480 progressive interlaced DVD rip.

Both Dione DV and Dione TV in VEAI 2.3 offer a pretty good quality and output a smooth 59.94 fps video.

Meanwhile, Video AI 3.3.3 using the same 2.3 models look a lot worse, like it’s missing half the resolution, the new IRIS model is a blurry pixelated mess and even the ‘Interlaced Progressive’ models, which output just 29.97 fps look much worse than VEAI 2.3 doing a 59.94 fps output… :person_facepalming:t2:

How can the quality be so backwards after so many years and updates??


I see similar problems. It seems Video AI 3.3.3. can’t do deinterlace at all while the previous version 2.3 does it well.


It’s so baffling that if anybody would tell you “Look, this is the new improved result of Video AI 4.0!”… about the 2.3 images you would say “Yeah, of course those look much improved!” :person_facepalming:t2:

I’m still new to Topaz Video (and only have lived with v3.x so far) and after a few experiments I found that Handbrake (ffmpeg inside, etc) was still more reliable for doing the first-pass extraction of MPEG-2 media. Specifically I trust it (a lot these days) for handling things like co-sited versus interstitial sampling, and 3D handling of vertical motion (bwdif filters and anything via the BBC R&D folks).

The fact is 4:2:0 is really hard to understand and then you add interlacing to it and your head explodes — and I really doubt anybody who was brought up on MP3 and XviD and learned about video in the torrent era, or in the 2020s, knows what they don’t know, about the painful stuff about SDTV. Even the latest 8K displays (especially when you don’t use Sony Master Series) will take chroma data and just do (chroma) pixel-doubling like it’s 1993.

To this day I still don’t understand some of the vertical staircasing that goes on in 720x576 sources (and I’m sure 720x480 is the same), where horizontal resolution seems artificially sharpened at the source; and yet depending how I process it — because, the very initial upscale is the most sensitive — I can end up with an authentic result or not. “Decisions Were Made”, long long ago, and if we ignore them the past will end up looking like dirt.

I’m really super awful keen that we don’t eff-up the source before we ask AI to clean it again.

Like, there’s a place for automation. And then there’s a place for yelling at the human “STOP WALKING ON THE CARPET WITH DIRTY SHOES” before you give it to the dry-cleaner.

Like, for instance, chroma dot crawl? You could ask AI to clean it. Or you could do a 4-frame / 8-field motion filter and just use the average image to remove it completely. Now clearly that’s a naive approach (and might be different for NTSC not PAL) but you retake 100% detail on a 100% static image. For any motion yeah, you have to be less naive — the number-crunching maths was all done by other people already though, e.g. the PAL Transform Decoder.

I haven’t tested in a few versions, but when 3 came out, for some of the Artemis models, it would select a bigger resolution AI model than version 2. That might be what you’re seeing here.
VEAI Version 2.6.4 DVD input to 1920x1080 output. Model: Artemis High Quality.
Try the same source and settings in TVAI 3.2, and you get the output of VEAI Version 2.6.4 with the same settings but as if you had changed the output to 4k and resized it down to FHD.
Sum-up for Artemis HQ with DVD input:
TVAI 2X scale = VEAI 4X scale

I found how to change the command to fix it, so I stopped checking to see if Topaz ever fixed it.

None of these were Artemis and all the resolutions were the same, manually set up by me each time: 720x480 source input to 1920x1080 output.

Mmmm… sorry, but what does any of that have to do with the topic of the post?

Can you please share the input file you are working with here? The screenshot, although helpful, are a bit difficult to compare since they are not side by side and the Iris model takes away from being able to compare the Dione shots.

You can securely submit your files(s) to our Dropbox using the link below. Please be sure to send me a note to let me know you sent something.

Submit File to Dropbox

If it can happen on some of the Artemis models, what’s to stop it from happening on some or all of the Dione models?

Here is the video with that area for anybody that may want to do their own tests: Topaz_Test_01.mkv - Google Drive

@bombayetas Out of curiosity, I noticed you mention v2.3 of VEAI, how is it compared using VEAI v2.6.4 (latest release of that line product)?
do you get the same “good” results with 2.64 as you getting with 2.3? or 2.3 in your test(s) outperforms 2.64?

Haven’t tried with 2.6.4, but my guess is they should be the same? I preserved a copy of 2.3 because it’s the last one that supported command line.

1 Like

Here’s a new comp sheet with only Dione models and 2x resolution, and what I think are the most interesting points: https://i.postimg.cc/RVRnZc5p/Comparision-Dione-x2.png

  1. The logo in the box: One part is the kanji letter in the center, which is clearly well preserved in 2.3 with both Dione DV and Dione TV and totally botched in 3.3.3.
    And the other important aspect is the shapes surrounding the kanji, the circle inside a square, with some lines in the upper area. Those 3 elements are also clearly a lot worse in 3.3.3 than in 2.3: The circle doesn’t maintain its shape (specially on the right side), the square shape is blurred and its corners fuzzy and almost disconnected.
    The horizontal lines in the upper part are almost totally lost in 3.3.3, as if only half of the image was processed.
    For further reference: here’s what this looks like in another part of the video where the camera is static, so there is no interlaced lines: https://i.postimg.cc/gzChSzHs/image.png

  2. Black line between light gray: straight with the shape well preserved in both 2.3 models, totally “half res” looking in 3.3.3

  3. Black strip on the cardboard box: clean and perfectly antialiased in 2.3… pixelated, blurred and showing halos in 3.3.3

  4. Dark area between boxes and boxes edges: clean and uniform with straight edges in 2.3. Noise and blurred edges in 3.3.3

Also worth noting is that I always liked a lot how VEAI 2.x could output a very smooth 59.94 fps video file from an interlaced source, thus improving not only the resolution and quality of the image, but also the motion smoothness.

Now in 3.3.x even the options that don’t smooth the motion and duplicate the fps look much worse…


And then IRIS: https://i.postimg.cc/k4Qghm7Y/Comparision-IRIS-x2.png

which is blurry and pixelated “half res” looking in ‘Interlaced’ mode or just totally blurry in ‘Interlaced Progressive’ mode; both losing all the fine details, as in the logo.


The image sequence in the zip file has been butchered. No one will be able to draw any meaningful conclusions from testing that.

Fixed: Topaz_Test_01.mkv - Google Drive

it’s a split of a sec movie, can’t see anything.
can you upload a 5 sec movie of your original and your results from all testing (Dione DV/TV mainly)
that way we can do a deep compare and testing ourselves.

Here’s a capture of my output which was a 480p to 1080p upscale in three Dione models from TVAI 3.3.4 and exported in ProRes:

Looking a bit different to yours.

1 Like