Topaz Video AI 7.2.0.1b

Did you ever figure this out it’s doing same to me rtx 4090 anon collection on. No option for local rendering.

There was a tip posted I think in the last beta thread. You can force it to quick export by typing “CTRL+E”. I think that was it. If that doesn’t work, check the previous beta thread. I had the same issue in the last beta. The export button was not there.

Thanks for the response. I did try the Ctrl+E method doesn’t work for me unfortunately.

It started working when Topaz fixed their end.

Is your output res set to 1080 or less?

Hello there, I think it is the same issue as Starlight Mini, it depends on your input resolution. Past a certain height (I suspect it is PAL 570i/p); you don’t get the “minimum” option anymore. Right now I am upscaling an old clip in 320x240 and the resolution slider for Starlight Sharp offers me the “Minimum” option at 1280x960.

This is not directly related to Starlight Sharp, but to the interface revamp : where have the codec options gone ?? In the new interface, it seems you cannot change on the fly per video the image and audio codecs and the container (beside the default choices in the Preferences menu).

As many other users, I really prefer the previous interface :slight_smile:

3 Likes

I noticed some changes not mentioned in the release notes regarding memory use.

Seems this release has changed the GB VRAM breakpoints. It used to be that 24GB cards OOMed if they specified 24 GB to the tvai runner.exe program. That no longer happens.

The new SLM breakpoints seem to be (tested on a dedicated GPU without display)

SLM
≤ 11GB uses 7GB
≤ 23GB uses 9GB
≥ 24GB uses 15GB

SL-Sharp seems to be constant at 9.4 GB for 4x upscale on my test clip regardless of what max-memory is set to.

What should we make of this, that SL-Sharp uses some fixed size tiling scheme like the older models do, and that SLM does not?

Second, did you make any trade-offs in order to halve the SLM memory usage, such as reducing the temporal aperture (batch size / range of frames)?
(which would reduce temporal cohesion I’d assume)

1 Like

About temperature, SL Sharp makes my GPU heat much more than SL Mini. A 6000 PRO Blackwell. Max 77° with SL Mini, Max 85° with SL Sharp. I had to turn on MSI Afterburner to lower the temperature

TADA !

I’ve played around with a few different media types trying to find where Starlight Sharp works best, and I haven’t found it yet.

  • Starlight Sharp on a high quality, medium/low resolution video (540p) seems to produce a over sharpened result which I’m not a fan of. However it did seem to be more temporally stable than Starlight Mini in my test.
    • I mentioned this in a previous comment.
  • Starlight Sharp on a highish quality, medium/low resolution video (E.g. DVDs) produces a mix of interesting upscales (E.g. Human faces upscale alright), while other areas (clothes) look low quality with excessive sharpening applied.
    • I mentioned this in a previous comment.
  • Starlight Sharp on low quality, low resolution video (E.g. A highish quality music video (with a lot of cuts) that YouTube then processed down to 360p) doesn’t work well. The lack of detail due to low resolutions and low bit rates in the original media results in all upscaling models having to do a lot of guessing. And Starlight sharp seems to be doing a worse job than Starlight Mini in the specific video I tested:
    • Issues caused by compression artifacts have more impact on the output of Starlight Sharp than Starlight Mini. E.g. In the images below, you can kind of make out where the compression blocks in the original media was with Starlight Sharp, but not Starlight Mini.
    • Some parts of the video have very little detail (E.g. A small face, tree leaves in the background, etc), so Starlight Sharp hallucinates a lot of detail that it can’t keep temporally consistent (Even with slow camera movements). Starlight Mini does a much better job, adding some detail, but not too much, allowing it to keep it much more temporally consistent.
      • I understand this is a issue with upscalers at the moment. Adding more detail makes it hard to keep temporarily consistent
    • Yet at the same time, some parts of Starlight Sharp are softer or blobier than Starlight Mini (E.g. On a close up on a person’s face, the mouth with Starlight Sharp is quite bad (two red lines for lips, and a dark line for the space between the lips), but Starlight mini produced something more natural. So it’s both adding too much detail in some cases (my previous point), and then missing stuff in other cases.
    • When Starlight sharp tries to produce a detailed texture (E.g. Tree leaves), there is a noticable pattern in it’s output that’s more pronounced than Starlight Mini.
    • Starlight sharp is very inconsistent when deciding if it wants to “enhance” motion blur/depth of field. And ideally it wouldn’t enhance motion blur/depth of field.
    • Starlight sharp doesn’t seem to like cross fades with faces in them. Although that’s probably a very difficult thing to work with.

Overall, I like that Starlight Sharp was more temporally consistent with higher quality footage than Starlight Mini, and it appears to be faster on my hardware that Starlight Mini, but that’s the only things I like about it at the moment.

The over sharpening on high quality media looks bad in my opinion.
The inconsistent results in the highish quality media looks bad in my opinion.
The poor handling of low quality media looks bad in my opinion. Although it’s possible this type of media was “too low quality”.

Maybe it will work well with medium quality content?

Sorry my input on Starlight Sharp has been spread out over at least two days at this point. It takes a long time for Starlight to process things, and I’m not willing to test using the cloud.

6 Likes

hang tight

What did you like so much about this Gaia v6 model?

No - Auto-renewal for your individual license. Pro is a subscription anyways, there isn’t an option there for you to change - and it won’t charge you of course.

It won’t - and that webpage will be deprecated soon, as these models are now available in Astra.

So actually the auto-renewal comment is unrelated but will make more sense with future announcements. Trust me there.

Well, it managed destroyed videos in much more detailed way, than v5, both CG and HQ variant. I always wondered, why it was pulled out - was Gaia v6 ‘too good’? Users were rather positive about it, if I recall correctly. :slight_smile:
I’d love to have this model back, to recreate some of my workloads I did back then, when it was possible to use this model. It’s not possible to use the old version of VEAI with this model available on RTX 3xxx cards, right?

No won’t bill you but stay tuned there. Yes, you’ll get Pro features for the duration of your active license here

1 Like

In the previous beta, I could use Ctrl-E. Now, exporting does not work. WHY?!
how can I test Starlight Sharp??

I had to uninstall the old beta version and delete all and any of the beta files on the computer. Afterwards I installed the new beta release and now it works fine. I think there was a tweak the other day on Topaz’ end to make it work also. Hope this helps.

1 Like

Interesting. Honestly we’re going to have to dig on this one - would love to see any examples. I don’t often hear this one mentioned.

1 Like

Yeah please try again after deleting the old Beta files and let us know if it works for you