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

Those are different models than the ones I used, which were the same models from VEAI 2.3 ( with the ‘Enable Previous Models versions’ option in TVAI’s 3.x Preferences)

You used the ‘Interlaced Progressive’ option, so the output would be 29.97 instead of 59.94 ( a lot smoother ).
Even then, the 2.3 results are better: cleaner, less noise, better at preserving shapes in the logo and the details in the kanji, less halos… and with a smoother 59.94 output.

image.png

It’s not made for playing like a normal video. It’s an uncompressed movie with 3 frames, focused on the area that has the key elements to compare each model. That’s the point and there is no need for anything more.

Anyway, any other source with similar characteristics will show the same results: VEAI 2.3 making a cleaner output, doing a better job at preserving details, and doing a perfect deinterlace even with a 59.94 output; while TVAI 3.x looks like it’s only processing half the resolution when using 59.94 output and even with the ‘Interlaced Progressive’ mode at 29.97 output the results are noticeably worse with any model, as stated in detail in the post above with the numbered key points.

Only Robust v4 is processed under interlaced progressive. DV v3 and TV v4 are processed under the interlaced option and you tested both of those in your comparison. My output comes out different to yours.

I will now go test the v2.3 equivalent models in v3.3.

Edit: Here are all 5 models:

Again my output is different to yours.

Hint: Don’t blindly trust the internal preview playback inside TVAI. View the actual files generated in an external media player like MPC-HC.

1 Like

So the preview playback inside TVAI is very different from the final result?? :person_facepalming:t2:Ludicrous…
If the settings you are adjusting and previewing differ so much from the final output… that’s like… a HUGE and ridiculous problem itself.
Anyway, I make a png output then and still 2.3 output is clearly better than 3.3.3 with the same models:

image.png

1 Like

Now 2.3 Dione DV v2 and Dione TV v3 compared to 3.3.3 Dione Dehalo v2 and 3.3.3 Dione Tv v4:

image.png

And 2.3 Dione DV v2 and Dione TV v3 compared to 3.3.3 Robust v4 and Robust Dehalo v2:

image.png

all these are captured from the png output, not the preview inside VEAI or TVAI.

Thank you all! We are looking into this

3 Likes

Can you please try square instead of original pixels? If you are still running into this, can you share a longer sample file?

Still running into this with any interlaced source: output quality in 3.x is just noticeably worse. Tried VEAI 2.6.4 an the output is exactly the same as 2.3.

Are you using square or original pixels? JIC you did not see my last reply;

Can you please try square instead of original pixels? If you are still running into this, can you share a longer sample file?

I used Square pixels. Difference is very noticeable in many details.
Here is another similar source: 720 x 480 interlaced. Dione 2x DV v3 in both 2.6.4 (left) and 3.3.5 (right): image.png

And here is the source. Full frame and 53 seconds: Topaz Interpolated Source and Comp.zip - Google Drive

Every single detail is worse in 3.x. It even has a lot more noticeable halos around the hands.

1 Like

Here is my Dion:DV i ran on your video.

Steps:

  1. Deinterlace (QTGMC / Slow Preset) + Conversion to PAR: 1x1 (Square Pixel) in Hybrid
  2. upscaled with Dione:DV using TVAI v3.3.6

How did this come out from your prospective? please check and revert back to me/us.

1 Like

Many parts look cleaner and smoother (and warmer tones) and many details are just gone, but I assume that is also due to the HVC1 compression, which is not good for a true comparision. Can you export to PNG or any other lossless output in the future? exact frame is 244,59.94, at time 00:00:04.071

2.6.4 Top - Hybrid+3.3.6 Bottom

Look at the monitor frame letters. In Hybrid+3.3.6 is just a blur. Same with the notebook. All the detail is gone: no lines in the page and the spine is also a blurry mess.

2 Likes

Any news on this problem?

Can you double-check the FFmpeg command from both exports to ensure that they are the same?

I think I already provided plenty of tests, analysis and sources for you to make your own tests.
I don’t think this has to do with the encoding part.

@ida.topazlabs
I’m pretty sure he’s reporting on this same issue that I and others have reported several times.
Topaz Video AI v3.0.0
Topaz Video AI v3.0.0
Topaz Video AI v3.0.2

I remember making a more detailed report, but couldn’t find it.

you mean this?

No. That one’s specific to Theia.
It had pictures of the same show, but was pointing out that a different model size gets used in TVAI than VEAI with the same settings. And that 1X, 2X and 4X all produce very different looking results. My conclusion was that TVAI picks the 4X model where VEAI picks the 2X model… At least with Artemis High Quality upscaling a DVD source to FHD.

Yes, that is what our thought process is with asking the OP for the comparison of the FFmpeg commands.

Thank you!