ONNX models to use with avs-mlrt

[kedautinh12]

About gain, maybe this model can help you "1x_Filmify4K_v2_325000_G"

[Reel.Deel]

Originally Posted by Dogway

Thanks for the conversions Reel.Deel!
Maybe HiNet can be converted? It had really astonishing results for scratches.
And just asking out of the blue, not sure if a film grain model as such exists, that is to add filmic grain, it can be an alternative to chickendream which is too slow. EDIT: yes, found 1x_Filmify4K_v2_325000_G but grain is not of great quality...
I also had 4x_Valar_v1 which is good for photo upscales it seems, but as with many of these it might require perfect artifact free downscaled content.

Also as a suggestion it might be interesting to try 2x_Waifaux-NL3-SuperLite_latest_G, as it might be a faster alternative to current solutions like w2xncnnvk.

No problem, I will add your other suggestions (edit: added). And HiNet cannot be converted (at least in ChaiNNer, who knows if there are other methods).

Originally Posted by Dogway

EDIT: Tested 1x_JPEG_60-80 and 1x-Kim2091_DeJpeg and got a system freeze, black screen and AvsPmod crash. Well anyway I compared them to Oyster and they paled against it. Since Oyster is so slow, it would be interesting to train a model based on it with my script from this post and probably a deblur model on it. Maybe if I have time in the future I can do such thing.

What were the dimensions of the image when trying 1x_JPEG_60-80 and 1x-Kim2091_DeJpeg? For me, I can't get most of the bigger models to run (black screen, driver crash) unless I use the tilesize and overlap options. With Kim2091_DeJpeg I had no trouble with but only used it on small picture. And I agree, I was not too impressed.

Training a model on Oyster does sound interesting. I read that Zarxrax used avs filters to train some of his models.

Originally Posted by anton_foy

Could you or someone convert 1x_ISO_denoise_v2?

Done.

[Dogway]

The same image from the linked post, 472x400 pixels. I will try the tiling option. I mean my GPU is not THAT bad : ) TBH I'm waiting for an affordable +10Gb card, be it this gen or the next one, the xx70 sitting on 8Gb for 8 years is as ridiculous as the current prices.

[poisondeathray]

Originally Posted by anton_foy

I had good results too with QTGMC's denoising in the past but it left some noise yet calmed down.

I wasn't referring to QTGMC for the purpose of denoising; I was referring to using single image filters on video such as 99% of the ones in that database - they can be prone to temporal inconsistencies. QTGMC inputtype=1 or 2 is one option that can be used after the single image filter(s) that can smooth over the consistencies, but as you know there are side effects to using QTGMC in progressive mode

Any link to the conversion software or does it cost?

chaiNNer is free. Reel.Deel posted instructions how to convert earlier.
https://github.com/chaiNNer-org/chaiNNer/releases

[Reel.Deel]

It's very easy to convert also. Load the required nodes, connect them and then hit run by either pressing the play button at the top or F5. Just make sure you've installed the dependencies.



Side note for Win7 users like me. Alpha v0.17.1 is the last version that works and you have to place the api-ms-win-core-path-l1-1-0.dll in the Python folder which appears after the first time starting the program.

[anton_foy]

Thanks Reel.Deel!

Quote:
Originally Posted by anton_foy View Post
I had good results too with QTGMC's denoising in the past but it left some noise yet calmed down.
I wasn't referring to QTGMC for the purpose of denoising; I was referring to using single image filters on video such as 99% of the ones in that database - they can be prone to temporal inconsistencies. QTGMC inputtype=1 or 2 is one option that can be used after the single image filter(s) that can smooth over the consistencies, but as you know there are side effects to using QTGMC in progressive mode

Quote:
Any link to the conversion software or does it cost?

chaiNNer is free. Reel.Deel posted instructions how to convert earlier.
https://github.com/chaiNNer-org/chaiNNer/releases

Aah okay I had too little knowledge of QTGMC I did not know about this mode. Thanks for the link, I will use it!

[DTL]

Originally Posted by poisondeathray

Prefilter can sometimes help, but yes chicken vs. egg.

Finally iterative solution for chicken and egg problem in motion compensating denoising - https://forum.doom9.org/showthread.p...52#post1984152

Now I think if it possible to replace either MAnalyse or MDegrain in that solution with some neural network processing from existing but keeping the iterative idea of checking each generation result with non-distorted input source.

In each iteration the algorithm refines both denoised clip and MVs field. But there is no current NN plugins capable of return MVs field in compatible with mvtools format ?

May be instead of MDegrain create an interleaved clip with MCompensated frames from input MVclip for each iteration and pass it to RIFE denoise ?

[Reel.Deel]

I'll respond here to not hijack Selur's thead.

Originally Posted by poisondeathray

No error message ? Just "hangs" ?

Yet other onnx conversions work ok on that same avs setup ?

Did you check that other applications which might be using GPU are closed ?

I closed everything, except VDub2 and it just hangs. I let it run for about 5 minuted and nothing. It says not responding and shows this:

Description:
A problem caused this program to stop interacting with Windows.

Problem signature:
Problem Event Name: AppHangB1
Application Name: VirtualDub64.exe
Application Version: 2.0.0.0
Application Timestamp: 5e73f48a
Hang Signature: 315e
Hang Type: 0
OS Version: 6.1.7601.2.1.0.256.48
Locale ID: 1033
Additional Hang Signature 1: 315e1db675843581067ed78f1c415427
Additional Hang Signature 2: 4ae5
Additional Hang Signature 3: 4ae5c6bd4a564e63c8d65dcb154a1aa2
Additional Hang Signature 4: 315e
Additional Hang Signature 5: 315e1db675843581067ed78f1c415427
Additional Hang Signature 6: 4ae5
Additional Hang Signature 7: 4ae5c6bd4a564e63c8d65dcb154a1aa2

I have not tried many models but all of Zarxrax's models work fine and also 1x_SwatKatsLite_360000_G.

Edit: It works now due to something very odd, this is my script:

FFVideoSource("HT1-1-RainbowsSample.m2v") # NTSC DVD
TFM(PP=0)
TDecimate()
ConvertToPlanarRGB()
o = last
ConvertBits(32)
mlrt_ncnn(network_path="...\2x_AnimeClassics_UltraLite_510K.onnx", builtin=false, tilesize_w=width/4, tilesize_h=height/4, overlap_w=8, overlap_h=8)
ConvertBits(8)
#Interleave(o,last)

When Interleave(o,last) is not commented out, it just hangs

[StvG]

Originally Posted by Reel.Deel

FFVideoSource("HT1-1-RainbowsSample.m2v") # NTSC DVD
TFM(PP=0)
TDecimate()
ConvertToPlanarRGB()
o = last
ConvertBits(32)
mlrt_ncnn(network_path="...\2x_AnimeClassics_UltraLite_510K.onnx", builtin=false, tilesize_w=width/4, tilesize_h=height/4, overlap_w=8, overlap_h=8)
ConvertBits(8)
#Interleave(o,last)

When Interleave(o,last) is not commented out, it just hangs

"o" and "last" have different dimensions.

[Selur]

"o" and "last" have different dimensions.

I agree, but it should complain about that and not crash,...

[Reel.Deel]

Originally Posted by StvG

"o" and "last" have different dimensions.

My bad, I completely overlooked that.

But there might be something else going on. Interleave always throws an error when the dimensions are different. Can anyone else reproduce this behavior?

[poisondeathray]

Originally Posted by Reel.Deel

But there might be something else going on. Interleave always throws an error when the dimensions are different. Can anyone else reproduce this behavior?

Just with a "normal" script ?

a=colorbars()
b=a.bicubicresize(a.width*2, a.height*2)

interleave(a,b)

Interleave: videos must be of the same size.

[Reel.Deel]

Originally Posted by poisondeathray

Just with a "normal" script ?

With a "normal" script it works fine. But when avs-mlrt is involved and the interleaved videos are different dimensions, it causes VDub2 to hang. At least for me. Can you try this script:

ColorBars(pixel_type="RGBPS")
o = last
mlrt_ncnn("2x_AnimeClassics_UltraLite_510K.onnx", builtin=false)
Interleave(o,last)
ConvertBits(8)

It works fine if o = last.PointResize(width*2, height*2).

[poisondeathray]

Originally Posted by Reel.Deel

Can you try this script:

ColorBars(pixel_type="RGBPS")
o = last
mlrt_ncnn("2x_AnimeClassics_UltraLite_510K.onnx", builtin=false)
Interleave(o,last)
ConvertBits(8)

I get the "Interleave: videos must be of the same size. " message in avspmod; but vdub2 hangs

[Reel.Deel]

Originally Posted by poisondeathray

I get the "Interleave: videos must be of the same size. " message in avspmod; but vdub2 hangs

Same. I guess it's a VDub2 issue.

[takla]

Fails silently with my AMD Radeon GPU...
Standalone vulkan ncnn converters do work. So the culprit must be the avisynth plugin.

[poisondeathray]

Originally Posted by takla

Fails silently with my AMD Radeon GPU...
Standalone vulkan ncnn converters do work. So the culprit must be the avisynth plugin.

Did you fulfill the requirements in the ReadMe ?

Other common "culprits" for some people's silent fails were non meeting the avs+ minimum version, and GPU memory (use tiles) . Neither current produces a useful error message, which could be an improvement

[takla]

Originally Posted by poisondeathray

Did you fulfill the requirements in the ReadMe ?

Other common "culprits" for some people's silent fails were non meeting the avs+ minimum version, and GPU memory (use tiles) . Neither current produces a useful error message, which could be an improvement

Lmao. I literally updated everything but avisynth+. That was the issue. Works now with r3958. Thank you

[Dogway]

I found new interesting models, not sure how feasible is it to convert to onnx, compatibility-wise:

Photos:
    Real ESRGAN 4x Plus
    Real ESRGAN General WDN 4x V3 (photos, paintings or anime illustrations)
    Latent Diffusion Super Resolution (LDSR) (SUPER slow though)
    SwinIR
    Swin2SR
    Remacri
Anime/LineArt Illustration Upscale:
    Real ESRGAN 4x Plus Anime6B

Also sometimes when refreshing the filter in AvsPmod, all I get is a pink screen. Looks like some out of memory issues, even if I use tiles. Need to restart AvsPmod.

[Reel.Deel]

Originally Posted by Dogway

I found new interesting models, not sure how feasible is it to convert to onnx, compatibility-wise:

Photos:
    Real ESRGAN 4x Plus
    Real ESRGAN General WDN 4x V3 (photos, paintings or anime illustrations)
    Latent Diffusion Super Resolution (LDSR) (SUPER slow though)
    SwinIR
    Swin2SR
    Remacri
Anime/LineArt Illustration Upscale:
    Real ESRGAN 4x Plus Anime6B

Do you have a link to Real ESRGAN 4x Plus and LDSR pth models?

I did not test them, but so far these models convert to to onnx without any problems:

• realesr-general-wdn-x4v3.pth : https://github.com/xinntao/Real-ESRG...ses/tag/v0.3.0
  
• RealESRGAN_x4plus_anime_6B.pth : https://github.com/xinntao/Real-ESRG...anime_model.md
  
• Remacri : https://upscale.wiki/wiki/Model_Data...iversal_Models
  
• SwinIR : https://github.com/JingyunLiang/SwinIR/releases

What specific SwinIR model you want converted, or all?

With the Swin2SR models I get an error when trying to covert:

An error occurred in a Convert To ONNX node:

Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument tensors in method wrapper___cat)

Input values (partial):
• Data Type: fp32

Not sure what that means.