Topaz Video AI v3.4.1

Hello everyone,

A new release of Topaz Video AI is now available.

Released August 22nd, 2023


Changelog from 3.4.0

  • Added second enhancement
  • Added new user guidance
  • Added preference to turn off new user guidance
  • Video player playback and trim functionality is more reliable and accurate
  • Improved dynamic bitrate usability
  • Improved tooltips

downloadmodels.bat not work

Hello I’m japanese beta tester
So my English is not so good. sorry

I download TVAI 3.4.1 and installed
Allway I use downloadmodels.bat
TVAI 3.4.0 is ok
But 3.4.1 causes error.
Ofcouse I copied downloadmodels.bat to TVAI 3.4.1 program folder.
not exist videoai.dll is error messgage
install trouble? or Bug?

Now I copied videoai.dll from other TVAI Beta program folder
Then downloadmodels.bat started

I assume this is the second pass enhancment from the 3.4.1 beta? If so, it’s missing from this release.

Hello there. The “second enhancement” function needs to be activated from Preferences.


Thanks for clarifying that. Topaz Labs really should of included a note about that in the release notes.


Edit: This has been answered below

Compared to the beta, there is no “Intermediate Resolution” option. Is the removal of this option intentional? The removal of this option makes the second enhancement feature unusable for some of the use cases I have for it.

And this parameter is also activated from Preferences. :))))
What I don’t understand is how “Intermediate Resolution” works when set to auto.

1 Like

Thanks, I missed it. Sorry for constantly commenting about non-issues, it’s just not clearly explained in the release notes.

TVAI will automatically pick what intermediate resolution to use. From a quick investigation here’s some notes.

If the output resolution is greater than 0x and less than 4x, then the intermediate resolution will be 1x. 
If the output resolution is greater than 4x and less than 8x, then the intermediate resolution will be 2x.
If the output resolution is 8x, then the intermediate resolution will be 4x.

The algorithm might be a little more complex than this (E.G. The auto algorithm may chose different cut off points depending on the input resolution), but that’s what I could identify from quick tests.

1 Like

when iris models uses TensorRT models on NVIDIA machines?


Thank you for the clarifications…

Is it just me or is it not fixed? :face_with_head_bandage:
Find 4 differences…

1 Like

Just out of curiosity… is there any way TVAI can utilize the new DLSS 3.5 AI stuff from NVIDIA along with your own stuff? Would be interesting to see if the NVIDIA stuff could mesh with TVAI and further enhance stuff, or if it would just clash like crazy.

1 Like

Also, updating your GPU drivers while TVAI is open, not in use, not cooking, still will crash the whole program.

That’s why it is generally a good idea to close all running programs when installing any kind of driver (or most applications). Reinstalling video drivers also crashes iCue (at least if you are using a Nexus or LCD pump cover). Personally, I reboot after installing updated drivers, YMMV.

1 Like

Oh for SURE!! GPU updates do not crash my icue, but I have seen plenty of apps that have error traps so at least it is not a hard crash. This 100% is not a “problem” but just kinda letting them know and all.

I had high hopes for the “second enhancement” feature.

I am restoring interlaced video
and I wanted to use the AI enhancements of Proteus, right after the de-interlacing pass of Dione.

It works as expected,
but I am noticing that if I also upscale the video’s dimensions,
I get unwanted sharpen-related artifacts.

I think that you need to look into the upscaling step.
I think Topaz needs to resize the deinterlaced video without attempting to enhance it, and then apply enhancements to the resized video.

For example what I am doing now is:

  1. Deinterlace the video with DioneTV and NOT resize it inside Topaz.

  2. Take the deinterlaced video and upscale it using Handbrake.

  3. Use Proteus to denoise and enhance the upscaled video.

This way I don’t get sharpen artifacts and It maintains much more original detail from the original video.

So unfortunately for me the second pass for the moment isn’t saving me any time as I still have to do all these steps manually.


New update causes the program to crash when running.

1 Like

I confirm, scaling in the second pass is not performed. If resolution control is performed in the first pass, for example, 1x (original), then if 2x is set in the second pass, then the output video is now 1x, although it should be 2x. Similarly, if the first pass is set to 2x, and the second pass is set to 2x, then the output video should be 4x. In this case, it turns out 2x.

If in the program the size of the output video must match the resolution control, then the priority of the intermediate resolution must be added in the first pass, and the second pass will correspond to the set resolution control for the output video and be automatic and increase or decrease the video according to the set output sizes.

I don’t even know what the programmers conceived, either the first option or the second, but for the second option, you need to specify it directly in the header (Intermediate Resolution for the first pass), otherwise, now it is very confusing, and then you need to rethink all the settings in order to make the right decision.

I don’t even know why the programmers went down such a difficult path, maybe it’s unique, but it’s too non-standard for the average user, who is used to simply adding various plugins one by one in other programs, knowing what everyone does and for what and in what order. I don’t understand why they are still trying to pull deinterlaced models, although this is already in FFMPEG, you just need to send commands, and check the boxes for the type of deinterlacing in the menu. See the documentation that is available with this program (FFmpeg Filters Documentation) C:\Program Files\Topaz Labs LLC\Topaz Video AI\share\doc\ffmpeg

1 Like

Is what not fixed? Can you clearly state what the issue is?

offset by 2 frames in this example, can’t you see? At the same time, video with a constant frame rate and without any speed filters. In version 3.4.0. this seems to have been fixed for video with a constant frame rate, although it remained for a variable frame rate, now it’s all back to frame desynchronization again.
To fine-tune, you just need a perfect comparison for identical frames in order to correctly set up all the filters in the model, here the frame desync in the preview is simply terrible and also jumps in the first window if there is a doubling of frames in the second. For frame interpolation, this is somehow understandable and you can close your eyes to comparing frames, but when the frames are constant, then the mismatch in the comparison is simply not acceptable.