Topaz Video AI v3.1 – faster exports, faster playback, and new Motion Deblur model

Topaz Video AI Roadmap
20

I have two 8 core machines - 1 AMD, 1 Intel. I have not got any version of this program to work on either machine after version 2.6 This one generated this message on the AMD machine

Last FFmpeg messages:

cur_dts is invalid st:0 (0) [init:0 i_done:0 finish:0] (this is harmless if it occurs once at the start per stream)

Last message repeated 2 times

nal_unit_type: 1(Coded slice of a non-IDR picture), nal_ref_idc: 2

cur_dts is invalid st:0 (0) [init:0 i_done:0 finish:0] (this is harmless if it occurs once at the start per stream)

Last message repeated 1 times

nal_unit_type: 1(Coded slice of a non-IDR picture), nal_ref_idc: 2

cur_dts is invalid st:0 (0) [init:0 i_done:0 finish:0] (this is harmless if it occurs once at the start per stream)

Last message repeated 2 times

nal_unit_type: 1(Coded slice of a non-IDR picture), nal_ref_idc: 2

cur_dts is invalid st:0 (0) [init:0 i_done:0 finish:0] (this is harmless if it occurs once at the start per stream)

nal_unit_type: 1(Coded slice of a non-IDR picture), nal_ref_idc: 2

cur_dts is invalid st:0 (0) [init:0 i_done:0 finish:0] (this is harmless if it occurs once at the start per stream)

Last message repeated 2 times

nal_unit_type: 1(Coded slice of a non-IDR picture), nal_ref_idc: 2

cur_dts is invalid st:0 (0) [init:0 i_done:0 finish:0] (this is harmless if it occurs once at the start per stream)

Last message repeated 1 times

nal_unit_type: 1(Coded slice of a non-IDR picture), nal_ref_idc: 2

cur_dts is invalid st:0 (0) [init:0 i_done:0 finish:0] (this is harmless if it occurs once at the start per stream)

Last message repeated 2 times

nal_unit_type: 1(Coded slice of a non-IDR picture), nal_ref_idc: 2

21

Confirm - NICE speed increase (GTX 3080)

BUT still No improvement for CG
Any advance anyone? I still have to use v3.6.4 for CG

CG
CG1473×657 123 KB

22

Trying out the stabilization + deblur option and it is definitely much faster in this model. Hoping the stabilization is better than previous versions which were a hit and miss with my attempts. Thanks guys for working on processing speeds.

23

Mine made me log back in, but I had no problem with my license.

24

Ok, 1st impressions on 3.1.
As promised I am seeing a HUGE speed increase on my 3080. The video I had going was a 4k video and was hitting about .8 fps, I canceled it since it was only 1/3rd of the way through, screenshot my settings, and did them EXACTLY like I had it, pulling 2.3 fps now.

The preview is just as fast. The video looks the same as 3.0.12, MAYBE a hair less deblurred. Now speaking of deblur, I tried the new motion deblur, and with my other deblur turned off and no sharpening, it made people look like plastic with crazy artifacts everywhere and pointy noses. Not sure if it is just this video, or it will be all of them. I will wait to see what others say, but it looked BAD BAD.

Everything else in this release I can see is 3.0.12, just MUCH faster. Well done guys!

1 Like
25

When it says:

does that mean we lose the ability to render with that model at all? Isn’t that the more accurate frame interpolation model? So if we ‘upgrade’ to this version we lose the option for that that we had in the 3.0.5 version? So for accurate frame interpolation it wouldn’t be an upgrade really?

Or does it let us keep the installation of v3.0.5 and have v3.1 installed too (both on the same pc)? I assume it will ask us when installing?

I assume this version doesn’t add all the ProRes formats for the PC version either (including ProRes 4444) or the .json files will need re-editing for that to try to get some of those working on the PC since it doesn’t list anything about ProRes in the changes (though ProRes was listed in the advert for Video AI previously).

1 Like
26

It appears the program is still not able to account for screen scaling. I have 1 4k monitor set to 125% and 1 1440p set to 100% and when I try to maximize the window on the 4k monitor it goes off the screen. Also preview sizes are not accurate. 100% is actually 125% etc.

The old 2.6 version didn’t have the maximize issue but did have the same preview issue.

Also the program always opens on my main monitor even when it was not on the main monitor when it was closed.

3 Likes
27

Hi Richardcrawford0,

I’m a newbie to Topaz (purchased 10/29/22) and initially thought I’d also have to purchase a new license after upgrading to Topaz 3.1 due to the screen it displayed after upgrade.

Then I logged on and 3.1 was licensed from my original purchase from 10/29/22. :slightly_smiling_face:

Don’t know if this may be the case with you and some that replied to your post, but thought I’d mention this just in case if it applies to your situation and people that replied to your post.

Apologies if this is not applicable to your situation.

Mike

28

I updated it and it didn’t show an option to keep the old version and it has removed the Apollo model. So thats seems to be a downgrade in terms of frame interpolation as that Apollo model was added to give increased quality of interpolation - especially non-linear interpolation (better than the Chronos models). It’s shame that’s been downgraded in what is otherwise meant to be an upgrade without it giving an option to keep the existing version when installing the new version. Maybe I could have changed it to a different folder for it to keep the older version too (it did show a folder option when I pressed the back button which it didn’t seem to show originally but didn’t give options for “keep current version” in the installation menu).

Is there any info on when that Apollo model that’s now been removed will be added again?

Also didn’t there used to be more than 2 Chronos models? It shows just Chronos and Chronos Fast - were there not more Chronos models than that before?

I also don’t see a ProRes 4444 output option. Hopefully that will be added one day in the PC version (and without having to edit json files).

Also in video enhance AI it used to have an option to quickly compare 3 models to the original one upcaled. There doesn’t seem to be a quick “compare” type option in this. It would help if there was I think.

29

Also is the new “motion deblur” option using Themis before the interpolation option in the chain? If so, is that helping with the interpolation (if interpolation was wanted by the user) or potentially making it worse?

ie. in a motion blurred frame the motion blur contains where things were at every time point when the shutter was open. By removing a lot of the motion blur from the frame before the interpolator sees it aren’t you giving the interpolating model less time info for it to work with?

eg. if time points 1, 2 and 3 were blended together into 1 frame (showing motion blur) it might be that the frame interpolation can see that the object has moved through the different points where the motion blur was. If the motion blur is removed before the interpolator sees it isn’t it going to have a harder time knowing what points an object passed through during that frame while the shutter was open (because with less blur it’s like a much shorter shutter time)?

If the motion blur would help with the interpolation and deblurring isn’t after the interpolation then maybe the place in the chain could be changed. Or maybe the user could have the option of where it is placed in the processing chain (if there are advantages and disadvantages for both positions in the chain - before and after the interpolation).

Another option would be if one AI model could do both de-blurring and interpolation (eg. like Apollo) in one go (where both are wanted by the user) - it might give better results that way.

1 Like
30

Nice !!! My 3080 is gonna be happy.
Thank you for the improvements :slight_smile:

31

The Gaia CG model stopped working properly quite some time ago. The last time that model functioned well, and actually worked really well on CG, was around 1.5.3, after that, there were changes made that caused it to mostly break for what it was good at and Topaz has not been interested in touching the Gaia models in a very long time.

It won’t help with the new features, and speed improvements etc, but if you have the chance to try it in 1.5.3 you may like the results better, otherwise with Anime, Proteus usually works well on it with tweaking the settings manually.

6 Likes
32

3.0.12 isn’t affected either

33

You can download 3.0.12 from here: Topaz Video AI v3.0.12

34

Incredible job. A work that used to take more than 12 hours now lasts 7 hours. I will see the final result tomorrow and tell you how good it was, but I didn´t think that the improvement was going to be so dramatic! Thank you :heart:

Edit: Nah, I got happy too fast. It took the same :frowning:

Hardware doesn´t help in my case.

35

I might be blind, but I can’t find it. Where did you find it?

36

Never mind, found it too :rofl:

37

Thank you i already found it. Also got a 30$ discount link yesterday , which makes the renew tempting :sweat_smile:

38

I must be on a completely different planet from others. The 3.1 update is CONSIDERABLY SLOWER here. I downloaded the update to try before purchasing the license renewal.

I’m using VAI mostly to deinterlace and denoise old family DV and VHS footage, occasionally upscaling. On my main computer, W10, I7-4770k, 16gb with GTX-1660 Super, on V2 I can process at .11 sec per frame. With V3.0, .13 secs per frame, and now with 3.1, its .5 sec per frame - almost 4x SLOWER. What the heck? I’ve tried several files with different CODECS, and various settings in the software to no avail. I loaded it onto an older machine with a GTX-1650, and the same speed degradation happens there as well. Really disappointing.

Any ideas? Yes, it would be nice to have newer hardware, but it’s worked reasonably well enough until now, for my needs, on what I have.

3 Likes
39

It’s exporting significantly faster, but now it keeps freezing completely as it gets near the end.

1 Like