ONNX models to use with avs-mlrt

[anton_foy]

Thanks I tried both of your suggestions but it is impossibly slow and I tried Vapoursynth vs-scunet on the same clip and 30 frames would take over 1 hour. So I guess this processing UHD with these models is out of the question for now.

[kedautinh12]

Waiting for cuda ver
https://github.com/Asd-g/avs-mlrt/is...ent-1771187328

[anton_foy]

@kedautinh12
Thanks for the info good news.
Something I am wondering about is the num_streams param. Where to find the specification about this for my GPU? May this speed up if I put this to max for my card?

[Selur]

The number of stream depends on the resolution of your input and the VRAM that is used with model X and 1 stream.
If you use a model with one stream and only a fraction of your VRAM is used, you can run multiple streams in parallel to speed things up.
=> increasing the streams will cost VRAM, but can speed things up
So if either the input resolution or the used model changes, the VRAM usage and thus the amount of streams your card can handle can change.

[anton_foy]

Originally Posted by Selur

The number of stream depends on the resolution of your input and the VRAM that is used with model X and 1 stream.
If you use a model with one stream and only a fraction of your VRAM is used, you can run multiple streams in parallel to speed things up.
=> increasing the streams will cost VRAM, but can speed things up
So if either the input resolution or the used model changes, the VRAM usage and thus the amount of streams your card can handle can change.

Thanks I will play around with this.

[kedautinh12]

Would like to see SwinIR models:
https://github.com/JingyunLiang/SwinIR/releases
https://openmodeldb.info/?q=SwinIR

[Dogway]

Yes, me too. I tried in vapoursynth and SwinIR didn't work for me there either. Looking for a solid photo upscaler.

[kedautinh12]

For Vapoursynth, why you don't ask HolyWu?
https://github.com/HolyWu/vs-swinir/...pen+is%3Aissue

[Selur]

https://github.com/HolyWu/vs-swinir works fine here,...

[Dogway]

Ok, I thought the SwinIR models in pth format were supposed to work. Thanks both.

[kedautinh12]

Would like to see CodeFormer models
https://github.com/sczhou/CodeFormer/releases

[Selur]

Assuming some folks here use those models on normal content, how do you deal with temporal inconsistencies?
(like this: https://forum.videohelp.com/attachme...73592/Face.mp4)
Personally, I use QTGMC with input type 3 to lessen this, but I was wondering what others do.

Cu Selur

[kedautinh12]

Would like to see UniScale models
https://openmodeldb.info/?q=Uniscale

[Selur]

You can convert those with chaiNNer (https://github.com/chaiNNer-org/chaiNNer),... (these are also included in the Hybrid onnx models collection)

[kedautinh12]

Would like to see SRFormer models
https://openmodeldb.info/?q=SRFormer

[Selur]

Would like to see UniScale models

Those seem to work fine here with vs-mlrt. (they are rather slow)

Would like to see SRFormer models

Best post such request over in the github issue tracker.

Cu Selur

[kedautinh12]

I can see it in other model. They use srformer to base for trained on
https://github.com/Phhofm/models

[kedautinh12]

Would like to see Cutie models
https://github.com/hkchengrex/Cutie?tab=readme-ov-file

[takla]

Actually it runs "ok" as long as you're not using any of the broken models from the dropbox link (I haven't tested them all, but 4 of 4 gave an error)
Should be fine if converted manually with "opset 17"

[takla]

Here are some models which I've converted to onnx (Mostly for Cartoons/Anime)
I can vouche for all of them working (tested on AMD GPU)
Original sources can be found here

2x_AniScale2S_Compact_i8_60K
Very good with BluRay sources

2x-LD-Anime-Compact
Amazing with high quality DVDs

2x_Garfieldjr_span48
Very Good for mediocre DVD quality

Example script:

LWLibavVideoSource("INPUT.mkv")
ConvertBits(32).ConvertToPlanarRGB()
mlrt_ort(network_path="C:\Program Files (x86)\AviSynth+\plugins64+\models2\2x-LD-Anime-Compact.onnx", builtin=false, provider="dml")