Video Enhance AI v1.7.1

Very interesting since OpenVINO is a Intel product and is supposed to work only with Intel processors, even on their website it states that it is Intel-only

I recently upgraded from 1080ti to RTX 3090
Render time for the Same file
1080ti 27 Days
3090 6 Days
The difference is profound and expedites the work flow

1 Like

It’s probably only “supported” for Intel platforms, and has optimisations for Intel specific accelerated code paths, but also has fallback code paths so final compiled code will still work on other platforms.

My understanding from one of the Intel forums is that it was written to be as fast as possible, but can be faster on Intel.

Thank you. I’m not super familiar with colour spaces but this does make sense to me. I’ll find a way to do it with PNG + Davinci Resolve.

Interesting Techpowerup shows why my 1080ti should never use the new FP16 models!

FP16 (half) performance: 177.2 GFLOPS (1:64)
FP32 (float) performance: 11.34 TFLOPS

They cut the performance by 64, to keep selling Quadro cards.

Hello,
could you give me an opinion on that result? I used Gaia HQ from 1.7.1.
https://gofile.io/d/jwZ0hn
There seem to be some weird lines in specific points, maybe a bad deinterlace?
Thanks

Here’s an interesting post about FP16 vs. FP32:
FP16, FP32 - what is it all about?

Basically newer GPUs like the RTX 2080 and RTX 3080 support 16bit floating point math that’s up to twice as fast as FP32. However there is a quallity hit if the AI algorithm is not adjusted to work with lower precision calculations.

VEAI 1.7.1 adds the FP16 option, for faster performance with newer GPUs. That’s great, but has anyone seen any visible decrease in upscaling quality if you have a FP16 capable GPU?

If there is a quality hit, sticking with an older FP32 GPU like the GTX 1080 Ti may be slower, but would output higher quality video. Perhaps there should be a check box in VEAI for the user to choose FP16 or FP32 (if there’s no quality hit, this idea is moot).

1 Like

Hi there.
I noticed that converting to 1080p Gaia HQ v5 results in 3x or more processing time than artemis Hq v8.
So far in my test I fail to see where all that processing goes. I think gaia looks best but it might be that given all the processing times it gets I will tend to see that.
Is it really 3x times better?
thx

I believe it is, if you check “use fast models”, it will use FP16. If left unchecked, it will use FP32
But it will have the same quality. 1.7.X does not use the same models (Gaia has been rewritten and that is the reason why it does not look as good as before) as 1.6.X and older models were visually better. I leave both 1.6.1 and 1.7.1 installed in my PC because of that.

1 Like

What is the source? DVD? I don’t have those discs… is the NTSC telecined or pure interlaced?

I have noticed that a lot of 1990’s - early 2000’s era discs are telecined DVD’s that decimate to 23.976 fps progressive. However, there are often still interlacing artifacts in the progressive frames. In many cases, they can be fixed by running QTGMC after detelecine/frame decimation, with InputType=2.

Here’s the AviSynth script Selur Hybrid generates using the above for such a disc:

ClearAutoloadDirs()
SetFilterMTMode("DEFAULT_MT_MODE", MT_MULTI_INSTANCE)
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\LoadDll.dll")
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\DGDecode.dll")
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\TIVTC.dll")
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\RgTools.dll")
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\masktools2.dll")
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\mvtools2.dll")
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\FFT3DFilter.dll")
LoadPlugin("C:\PROGRA~1\Hybrid\32bit\AVISYN~1\nnedi3.dll")
LoadDLL("C:\Program Files\Hybrid\32bit\avisynthPlugins\libfftw3f-3.dll")
Import("C:\Program Files\Hybrid\32bit\avisynthPlugins\QTGMC.avsi")
Import("C:\Program Files\Hybrid\32bit\avisynthPlugins\Zs_RF_Shared.avsi")
# loading source: E:\USER\wip\rip\muxed_vob.mpg
#  color sampling YV12@8, matrix: bt709, scantyp: telecine, luminance scale: limited
MPEG2Source(d2v="mpg_dacd0eda52458c8ab739aa4cb79b9a64_853323747.d2v")
# current resolution: 720x480
# deinterlacing
TFM(slow=2,mChroma=true)
TDecimate(cycleR=1,cycle=5,hybrid=1,vidDetect=3,conCycle=1,sdlim=0,nt=0,blockx=32,blocky=32,chroma=true,denoise=false,ssd=false,display=false)
AssumeFrameBased()
# filtering
# denoising filtering using QTGMC
QTGMC(Preset="Placebo", InputType=2, TR2=0, Sharpness=0.2, SourceMatch=3, Lossless=0, MatchPreset="Placebo", MatchPreset2="Placebo", ediThreads=2)
PreFetch(4)
# setting output fps to 23.976fps
AssumeFPS(24000,1001)
#  output: color sampling YV12@8, matrix: bt709, scantyp: progressive, luminance scale: limited
return last

That output .mp4 would then be run through VEAI. I just use a constant rate factor of 1.

Source Is a PAL DVD and checking on internet I can confirm it’s interlaced.
The two problems are: those artifacts which I supposed are caused from a bad deinterlacing and an unsharp image, I don’t know how to get rid of them.
I’ll try your script for that artifacts and later try to work on sharpness(the issue Is that the more they are distant the more they are unsharpened).
Thank you for your answer!

I believe you are reading that incorrectly, as I see no such choice. There is “Do not Download Fast Models” which refers to models other than the default 288x288 versions that ship with the installer. I think if you look at your log files you will find that FP16 is always used by a 20xx GPU whether that box is checked or not.

1 Like

If it’s a PAL DVD then you should not need to de-telecine. Just run QTGMC with InputType=2, or, if it looks better, you can try InputType=3. I’ve in some cases even just run it as a deinterlacer with InputType=0, it’s just a matter of what looks best.

Since it’s a PAL DVD you might want to consider changing the frame rate to 24 fps and slowing the audio down in an audio editor like audacity, since that should be the original speed.

my bad then, I was under the impresion that only fast models would use FP16. I’m primarily using 1.6.1 for my own projects (which usually takes several days) and 1.7.1 only for other people’s projects. My advice is unless you’re using RTX 3000 series or AMD, stick with 1.6.1

I’d like to add that my GTX1660 Super is only using FP16 models in 1.7.1 no matter the selection of "Do not Download Fast Models’. Just thought it might be useful for others.

1 Like

Bring the Win7 support back, please!
Vulkan is better than DirectX12 , it can support Win7+ and Linux , and then support Macos by MoltenVK

Vulkan ML(Machine Learning)

Visual Studio Info. from Microsoft:
[ For the Windows target platform, this specifies the version of the Windows SDK that your project requires. When you install a C++ Workload by using the Visual Studio installer, the required parts of the Windows SDK are also installed. If you have other Windows SDK versions on your computer, each version of the SDK tools that you have installed appears in the dropdown.

To target Windows 7 or Windows Vista, use the value 8.1, since Windows SDK 8.1 is backward compatible to those platforms. In addition, you should define the appropriate value for _WIN32_WINNT in targetver.h. For Windows 7, that’s 0x0601. See Modifying WINVER and _WIN32_WINNT. ]

[Visual Studio]# General Property Page (Project)
https://docs.microsoft.com/en-us/cpp/build/reference/general-property-page-project?view=msvc-160&viewFallbackFrom=vs-2019

Say again:
Vulkan can use in machine learning same as DirectX12 , but Vulkan can use on Win7 , Win7 has at least 20% market share , don’t drop Win7 support till the Win7 market share less than Win8.1 , please!
Also, please set the [ Windows SDK Version(Target Platform Version) ] to the value [ 8.1 ] before you build the new version software.

And then , if you can’t make the Vulkan ver. AI Engine in short term, please use the D3D12On7(The Direct3D12 runtime for Windows 7) now.
Nowadays, many people still using Win7 , Win10 is very bloated and RAM usage is too high.

D3D12ON7 sample:

D3D12ON7 Packget (7zip can unzip the nupkg file):
https://www.nuget.org/packages/Microsoft.Direct3D.D3D12On7

3 Likes

It would be better if VEAI ran on more platforms / different OS’s. I agree with the above!

2 Likes

Hello Chase
Thank you for the answer!
I’m usually running QTGMC with input type=0 to only deinterlace, I’ll try to use the other values and check if it’s better on that source.
About the frame rate I’m keeping It at 23.98, Is It a problem or Is It as good as 24 fps?
Thanks!

As I understand it, 23.976 is more of the “accepted” standard, but only because it will allow the video to playback on older CRT based monitors and other certain (mainly North America) equipment. On modern equipment it shouldn’t matter. But keep in mind audio will go out of sync if it isn’t re-timed correctly (i.e., to 23.976 vs 24).

From a PAL source audio signal Is for 24 fps normally? If It Is, It would be better to encode to 24 fps and avoid to increase the needed workflow, right?
Thanks!