Topaz Video AI 6.0

When it comes to pure Processing Performance then your right :slight_smile:

What i mean is the responsiveness of 2.6.4

2.6.4 was like a Naturaly Aspirated 4 Cylinder Engine, not so much power, but it reacts fast and has not much overload (at least not on the HD/SSD Side)

Newer Version are more like 6 Clyinders with a big Turbo Lag.

I see it as a Advantage that i can see in realtime (but not at full speed, so that i can spot AI Artefacts better :wink: ) what 2.6.4 does in the Preview, without stuttering.

With New Versions it’s more like that.

Click on a place in the timeline, click on Preview 240fps.

Video AI Starts after 1-2 Seconds, stuttering Preview Output.
Click on Preview Picture, Seek Bar snaps back to first processed Frame, Click Play…
-Repeat Start
Output Starts then stops at Frame X because Video AI is not done yet.
Wait !
When a few Frames more are done, playback resumes.
-Repeat Stop

The repeat part goes on and on, until Video AI is done.
After that i can watch the full preview in full speed (what i dont want, because it’s harder to spot short AI Artefacts on the skin & face)

1 Like

I agree on it all. :ok_hand:

And while v2.6.4 processing wasn’t saturating your CPU/GPU enough, you could easily run several processes concurrently, to offset that 1 negative aspect about it.

2 Likes

Yes your absolute right.

I realy realy hate it, when my complette system (Ryzen 7 2700, 32GB Ram, RTX 3060, 2TB SSD. 400GB Free) is blocked (just something simple as a Window Change takes up to 20 Seconds) when the newer VideoAI Versions “Finalize” the Output.
The wear leveling from the SSD has to work like crazy.

1 Like

I checked it out, and it yields results worse than TVAI.

5 Likes

I do this often in my video cutter software; TVAI upscale with two different models, and then overlay, I set opacity pending on szenes. Two or more models used/blend szene based into TVAI, would be a great thing!

But now is better stoping those expectations, first Topaz should do everything to fix their bugs! …and deliver a space-saving interface with all needed information, they can also go back to the old GUI :smirk:

2 Likes

I use VapourSynth myself, but same difference: they’re both very powerful frame servers.

I only know of Lancsoz upscaling filters (for best upscaling; the same TVAI uses in ffmpeg, btw). But yes, you bring up a good point: TVAI filters tend to do too much: you don’t always want a ‘Jack of all Trades’ all in 1 filter. Secondary Pass (if it worked) is a good start, but why not just multi-pass functionality? Where you can just endlessly press a + button, to add a new filter for the next pass?

image

5 Likes

It does. TVAI’s models, and flexibility, are, as of yet, unbeatable. But Neofootage AI definitely has potential.

2 Likes

Good news, in my case the “require login on each app start” bug looks like solved.
This was annoying in V 5.5.1 .

1 Like

I brought up the lack of preview render queues and stats in the alpha as well.

1 Like

That + a Button Multipass Idea is very nice :slight_smile: and would
be usefull, because Video AI is sometimes very unpredictable in the outcome, especialy with Interlaced Material and Dione.

Sometimes you throw in DVD9 Material, with high Detail, Low Blocking, and nice Sharpness, and use Dione DV, and the Outcome is full of weird Artefacts. Fine Lines are converted to a pattern that look like this char → ~ ←

Then u add second enchangement with Scale 2x, Artemis Aliasing and Moire, Recover Detail 100%, then the ~ patterns are gone, the contures are cleaned up, and finer, but some real detail & grain (which i want to keep. Pron has to be look good in Motion, not just on one specific Frame) is gone.

Then you compare that Output, with a Medium Setting QTGMC Output, and the QTGMC Output has more Original Detail, but less sharpness (especialy on Hairs) on the Outlines from a Person.

But all in all the QTGMC Output looks in Motion more alive and real (So it’s better for my taste)

The way how details on gradients on legs/arms are distributed, how grain, a low noise pattern, and details are stable (and dont come and go for a few frames like in Video AI) is just more harmonic. In germany we say “sieht aus wie aus einem Guß” which translates to something like “it looks like it was molded at once, from one piece of steel”

After you see how much time you wasted, you try Dione TV (without Second enchangent) and it looks amazing (for the most part, sometimes some weird AI Artefacts can happen) no crazy dehaloing, it just looks right !

Then you try a Source with something that has a little bit less Detail and Sharpness, you got the same Problems with Dione DV, and when you try Dione TV, it gets way worse.
Because for some reason Dione TV Dehalos sometimes the s*** out of the picture, and kill all the Details & Contures from the Picture.

Then sometimes when i throw in something from a DVD with a mushy washed out picture, and try Dione DV, the result is much better as anything that i can achive with QTGMC alone.
Then for some reason Dione DV creates nice looking fake detail everywhere, which just look nice :slight_smile:

But all in all: Most times i just lose Time when i try Video AI first for Interlaced Material, and QTGMC would be the better choice in the first place.

Because all what i need to do in VDub2 is change a little Setting, Press F5 and i will immidiately see the result.

1 Like

Yes, multi-pass would naturally have to come with some restrictions. Especially for filters that do frame interpolation, VFR (yes, on purpose), or, indeed, deinterlacing. A simple (internal) exclusion table would take care of that.

1 Like

God, I wish we could get the UI of v2.6.4 back!

6 Likes

YESSS, the frame duplicate interpolation/removal functions as intended! Now I’m really curious about the evolution of it! :slight_smile:

I like how the previewing is now very natural - just press PLAY and see. Preview files returned to their respective place - in the VEAI Workspaces - great! :slight_smile:
Now just the details, like ETA with pre-determined previews and the app activity notifications (downloading, preparing, erroring, as I just got one…) :slight_smile:

The error seems to have to do something with live render preview with Rhea v1 (or maybe my too excited clicking?) :stuck_out_tongue:
The crashing part is C:\Program Files\Topaz Labs LLC\TopazVideoAI3\ffmpeg.exe
, with 0xc0000409 BEX64 exception in ucrtbase.dll library. The app itself doesn’t crash at that moment.

30s ‘traditional’ RHEA v1 preview - interesting…

RHEA XL v1 ‘capitulates’ with 30s preview even sooner:

At this point, ffmpeg execution fails after a moment, it’s fairly random, when (Proteus v4 30s preview):
obraz

logsForSupport.zip (94.1 KB)
logsForSupport2.zip (161.3 KB)

EDIT: I’ll have to re-check this WITHOUT SD-AI working in the background… Commit charge went to the sky for some reason, while RAM wasn’t even fully utilized… :sweat_smile:


…and yes, it was the choked memory commit charge, which caused these crashes. Now I’m still running the TVAI v6 without the hitch, while testing previews extensively.

I’m also really glad, that TEMP folder wipe issue is also fixed. :slight_smile:

First impression of version 6.0.
Hurray, a long list of presets works. Okay, I’ve set some preset and… where the hell is the IVTC option for the source file? Okay, I’ve set it up, time for a test preview. I click preview and guess what? Crash! My first TVAI crash.
logsForSupport.zip (3,1 MB)

The perpetual license is nice.

1 Like

That’s why I don’t bother deinterlace my sources with TPAI.
I always use Hybrid, QTGMC with: Source Matching 2x refined; EZKeepGrain 1; Lossless before resharpening; Final temporal smoothing 0; EzDenoise off.
It keeps fine details like original noise and blocking for TPAI to take care of.

4 Likes

Two observations, not extensively tested or verified but this seems to be the situation:

After an export, if I return from the Export queue to Adjustments with the intent to fine tune my adjustments further, I find that all my parameters (Enable Parameters) in Adjustments have reset to their defaults. That’s a problem. It seems that if I hadn’t saved my Preset immediately before I ran my last Export (which granted is best-practice) so that I can reload my Preset, I’ve simply lost my latest parameter adjustments.

(It also seems that there is no longer a way to update/resave a Preset as one fine tunes it? Have to resave it as another Preset… and another… and another?)

Second observation is that the split-screen preview doesn’t align the two images over each other correctly if the original is SD and the output is HD.

2 Likes

You need to check ebay for some enterprise SATA SSDs. You can usually find some Samsung PM863A drives for a reasonable price. And even though they’re used, you’re not going to worry about wearing them out any time soon.

2 Likes

Yeah in the end you save many time if the source is from average quality or better, if you use QTGMC before you even consider to use Video AI for interlaced material.

But when Dione fits (which is rare for SD 4:3 Pron ^^) then the Result can be outstanding, totaly without Upscaling :slight_smile:

But even then, i Post Process the Video AI Output with AviSynth+ :smiley:

Just a few days ago Dione TV did such a outstanding Job at a DVD9 File with 640x480 (DAR) that i kept the Resolution, and just tweaked it a little bit with AviSynth.

With Upscaling to 1280x960 i had to use a CRF of 18 (and extreme Settings) with x264 to keep the grain and details, and had a File with a average Bitrate of 1200kb/s (9600kbps) The Result was just just slightly sharper (on my 19" 1280x1024 Display) compared to the 640x480 Output (which is Displayed with a normal Cubic Resizer in MPC-HC) which looked great (CRF 12) and took just half the space.

A Small Monitor with 1280 Width Native, can sometimes be a bless :smiley:

3 Likes

Thanks for the Tip :slight_smile:

SSD’s with MLC or SLC are in a totaly different league (got one 128GB MLC in my Second machine, that thing runs since 10 years like a beast. Condition is still at 99%)

1 Like