Video Enhance AI v1.7.0

In the test, I used the default configuration.

1 Like

awesome. have u tried 1.7.1. i just did my first upscales with it and its fast, and has the sharpeness of 1.6, they listened to us about how gaia looked soft. I will say however tho, artemis looks worse in 1.7.1. but gaia looks good. This is a conversion i just posted to youtube using 1.7.1 - YouTube

1 Like

I feel I wound rather wait to upgrade from 1.6.1 until all of the models have matured to the point where they all look good, aren’t too soft, and don’t leave artifacts. (Obviously not every model will look best on every source, but there shouldn’t be issues on a given source where there weren’t issues in the past.)

interesting, do you get good results by doing that in 1.7? How good is compared to 1.6.1?
I will try that after the current job is done
I actually didn’t like the result of GAIA-CG on 1.7 because it looked too smooth and didn’t seem to improve details and went back to 1.6.1. It takes twice the time (meaning 16 days now lol) but the final result is good.
Also, what is your workflow when doing that?
Like I usually go for
first 100% denoise/deblock at original resolution, than a second pass to enhance resolution (SD->HD or HD->UHD), a third pass if original file is lower than 720p (HD->UHD).

I don’t understand what you mean GAIA either HQ or CG is in the same V5 version in 1.7.0 or 1.7.1

I just tried the Aremis HQ v8 model. Compared with the 1.7.0 v7 model, the v8 model picture has become a bit cleaner. Maybe this model has not been trained in anime CG. I temporarily return to the 1.6.1 version and continue to use Gaia- CG model.

@Chase, You mentioned using TR2=0 in QTGMC. Your logic is sound, however in the real world, sources have noise, sometimes severe enough that retouching in Photoshop is needed. I had one letterboxed DVD movie that I edited 961 frames before sending to VEAI (this is AFTER using my QTGMC settings). I tried your settings, QTGMC(Preset=“Placebo”, InputType=0, TR2=0, Sharpness=0.2, SourceMatch=3, Lossless=0, MatchPreset=“Placebo”, MatchPreset2=“Placebo”, ediThreads=2).
…and the results were terrible compared to properly denoising without losing detail before sending to VEAI. With those settings, there was no increase in detail. There was actually less and lots of noise. The following settings are from 7 years of using QTGMC, studying the documentation and settings others have posted. This or a slight variation is what I use on DVD and VHS material (AG-1980 plus TBC) with excellent results: QTGMC( Preset=“Placebo”, NoisePreset=“Slower”, ChromaMotion=True, TrueMotion=True, TR0=2, TR1=2, TR2=3, Rep0=3, Rep2=3, NNSize=3, NNeurons=4, EdiQual=2, Precise=True, NoiseProcess=2, GrainRestore=0.3, NoiseRestore=0.01, Sigma=2.0, Denoiser=“dfttest”, DenoiseMC=True, NoiseTR=2, NoiseDeint=“Generate”, StabilizeNoise=True, EdiMode=“NNEDI3”, EdiMaxD=16, Sharpness=0.3, Smode=2, SLMode=0, SLRad=3, Sbb=3, MatchPreset=“Placebo”, MatchPreset2=“Placebo”, SourceMatch=3, MatchTR2=1, MatchEnhance=0.85, Lossless=2 ). I sometimes add back in a little noise, post-VEAI, with a MSU Noise Generator to keep it looking real (that’s a Virtualdub2 plugin).

Both versions use GAIA-HQ V5 though.

I’ll try your settings too. QTGMC has always given me excellent results, but Placebo looks overkill, how low does it take you to deinterlace content? A 1080i source deinterlaced with preset on slow is already killing me with 6 hours for 40 minutes of content :smiley:

I’ve been reading too, but not for 7 years, and so far, I’ve found it’s really hard to get people who are willing to share their FULL settings, like you just did. It’s more common for people to say “your settings are wrong” or “your results are terrible” but then either not say what to use, or just “allude” to what to use, or maybe share a few settings but not the complete string.

I’m using Selur Hybrid for the most part, and although the developer recently included some additional settings and controls for QTGMC, he didn’t include all of them (like “noise preset” for example), so I’ve been experimenting myself and the QTGMC setting I posted was the best I could come up with without losing detail. For example, I’ve found that using “Lossless” often re-introduces interlacing artifacts.

Hybrid does allow custom filters so I tried using your exact string, and the results look very good so far! Hybrid has a previewer, and I checked the “show settings” option for QTGMC, and I could see that many of the settings in your string were different than what defaults to “Placebo” and the other settings Hybrid has, and what I was using. QTGMC will mostly always default to using ff3d instead of dfftest for the denoiser for example.

I’m in the process of trying your string out and then upscaling with VEAI to see if I can get better results.

THANK YOU VERY MUCH for sharing your COMPLETE QTGMC string!

To be fair, most of the times QTGMC(preset=“slow”) will give you satisfying results, or a slower speed if you can afford it. I’ve experienced with a couple of fine tuning but I’ve never seen such a massive QTGMC call.

with my 16" mb pro the app crashes at launch. does anyone know why this is happening?

Process: Topaz Video Enhance AI [946]
Path: /Applications/Topaz Labs LLC/*/Topaz Video Enhance AI.app/Contents/MacOS/Topaz Video Enhance AI
Identifier: com.yourcompany.Topaz-Video-Enhance-AI
Version: 0
Code Type: X86-64 (Native)
Parent Process: ??? [1]
Responsible: Topaz Video Enhance AI [946]
User ID: 501

Date/Time: 2020-11-16 23:57:54.546 +0300
OS Version: Mac OS X 10.15.7 (19H15)
Report Version: 12
Bridge OS Version: 4.6 (17P6610)
Anonymous UUID: A6977279-C3F6-5162-879C-06C770E16225

Time Awake Since Boot: 280 seconds

System Integrity Protection: enabled

Crashed Thread: 0 Dispatch queue: com.apple.main-thread

Exception Type: EXC_BAD_ACCESS (SIGSEGV)
Exception Codes: KERN_INVALID_ADDRESS at 0x0000000000000008
Exception Note: EXC_CORPSE_NOTIFY

Termination Signal: Segmentation fault: 11
Termination Reason: Namespace SIGNAL, Code 0xb
Terminating Process: exc handler [946]

VM Regions Near 0x8:

__TEXT 0000000107cc2000-0000000107ddf000 [ 1140K] r-x/r-x SM=COW /Applications/Topaz Labs LLC/*/Topaz Video Enhance AI.app/Contents/MacOS/Topaz Video Enhance AI

Thread 0 Crashed:: Dispatch queue: com.apple.main-thread
0 org.qt-project.QtCore 0x0000000110b82ff5 QThread::start(QThread::Priority) + 37
1 com.yourcompany.Topaz-Video-Enhance-AI 0x0000000107cde8a9 TBackend::TBackend(QObject*) + 4537
2 com.yourcompany.Topaz-Video-Enhance-AI 0x0000000107cdb36a main + 1530
3 libdyld.dylib 0x00007fff6d35fcc9 start + 1

Thread 1:
0 libsystem_pthread.dylib 0x00007fff6d55fb68 start_wqthread + 0

Thread 2:
0 libsystem_pthread.dylib 0x00007fff6d55fb68 start_wqthread + 0

Thread 3:
0 libsystem_pthread.dylib 0x00007fff6d55fb68 start_wqthread + 0

Thread 4:: QQmlThread
0 libsystem_kernel.dylib 0x00007fff6d4a73d6 poll + 10
1 org.qt-project.QtCore 0x0000000110dba180 qt_safe_poll(pollfd*, unsigned int, timespec const*) + 608
2 org.qt-project.QtCore 0x0000000110dbb981 QEventDispatcherUNIX::processEvents(QFlagsQEventLoop::ProcessEventsFlag) + 849
3 org.qt-project.QtCore 0x0000000110d53e0f QEventLoop::exec(QFlagsQEventLoop::ProcessEventsFlag) + 431
4 org.qt-project.QtCore 0x0000000110b81a63 QThread::exec() + 131
5 org.qt-project.QtQml 0x0000000110893839 0x110601000 + 2697273
6 org.qt-project.QtCore 0x0000000110b82a79 0x110b60000 + 141945
7 libsystem_pthread.dylib 0x00007fff6d564109 _pthread_start + 148
8 libsystem_pthread.dylib 0x00007fff6d55fb8b thread_start + 15

Thread 0 crashed with X86 Thread State (64-bit):
rax: 0xf1db7aa3214200a0 rbx: 0x00007ffee7f3d9a0 rcx: 0x0000000107dd9610 rdx: 0x0000000000000013
rdi: 0x0000000000000000 rsi: 0x0000000000000007 rbp: 0x00007ffee7f3d6a0 rsp: 0x00007ffee7f3d600
r8: 0x0000000110f8f3b0 r9: 0x0000000000000000 r10: 0x00007ffee7f3dc48 r11: 0x0000000110b82fd0
r12: 0x0000000107de1950 r13: 0x0000000107dda615 r14: 0x0000000000000000 r15: 0x0000000000000007
rip: 0x0000000110b82ff5 rfl: 0x0000000000010206 cr2: 0x0000000000000008

Logical CPU: 10
Error Code: 0x00000004 (no mapping for user data read)
Trap Number: 14

Yes, I don’t know how some people here can even managed to process HD videos with placebo preset.
QTGMC was build for SD interlaced inputs. Even on my 16c/32t machine it simply don’t start on HD videos with the placebo preset and even on the slower preset I may have about 3fps at the best.

Here’s a comparison I did out of curiosity.

Placebo is @Brandon42 exact settings.
Slow is simply QTGMC(preset=“slow”)
Slow+MCTD is like Slow, but followed with a call to MCTemporalDenoise(settings=“low”, strength=50)
Slow+EZDenoise is QTGMC(preset=“slow”, EZDenoise=1.0)

I honestly think the placebo setting is the worst looking result (no disrespect) as it produces a blurry picture, on top on being extremely slow. Default slow is quite sharp and has good fidelity to the source. There is a faint noise on this source so I added two denoising solutions, EZDenoise works quite well and is about as fast as Slow. MCTemporalDenoise is very slow, but faster than placebo. The result of MCTD is comparable to placebo, albeit retaining a bit more detail. Results may vary depending on material, but you can make up your mind for yourself.

By the way, if you know where that picture comes from… don’t tell :wink:

lol I was going to ask but then maybe it’s better to let my mind think it is from any random japanese drama

So after trying your string on a 640x480 source it just isn’t looking as good as it does when using my string with no temporal denoising beforehand. Details are getting washed out and softened, as I expected they would. I’m using Gaia-CG in 1.6.1. I tried Gaia-HD and on the particular source I’m currently working with, Gaia-HD hardly did anything. Both with my setting, and even less with yours.

I guess it largely depends on the source and the model used. In my test it didn’t yield the same results. You mentioned VHS and maybe it helps with that. If you are saying you see a better improvement with your settings, on your sources, I don’t doubt you.

1 Like

самое главное сделайте Deinterlacing ! а потом уже все осталное ! сделайте пожалуйста!!!

DVD , DV, 1080i все приходится переделавать чтобы начать работать с Topaz Video Enhance AI ! потеря качества и времени! ну и конечно верните Gaia-CG и Gaia-HD с предыдущей версии !! что касается других алгоритмов то заметнее лучше и дело не в скорости а в эффективности!!

most importantly, do * * Deinterlacing** ! and then everything else ! please do!!!

DVD, DV, 1080i all have to be redone to start working with Topaz Video Enhance AI ! loss of quality and time! and of course, return Gaia-CG and Gaia-HD from the previous version !! as for other algorithms it is more noticeable better and it’s not about speed but efficiency!!

@plissandro Well, I guess you could say it is :smiley:

@Chase On a sidenote, my comparison involved a 1080i source. For SD material I usually go with a different setting for QTGMC that you might want to try, QTGMC(preset=“Slower”, SourceMatch=3). Usually retains better details. You might want to play with the Sharpness or TR2 flag if the result is too sharp for you (specifically I sometimes like to put them both at 0 if I run an independent denoiser like MCTD after).

As much as i know so far is, that audio support is not yet reliable.
If you need perfect audio, recommendation is:
Just enhance the video and mux audio back to the ready enhanced video.

Adobe Media Encoder ist good for such jobs.
And it perfectly exports jpg sequences from video, too.

MKVmerge is ideal for the task and it’s free

Not saying that Adobe’s app is any bad, just another suggestion

1 Like