vs-basicvsr

[Selur]

Only tested that it works with my A380, but never did any speed test since my RTX 4080 is clearly faster.

Cu Selur

[PatchWorKs]

Originally Posted by Selur

Only tested that it works with my A380, but never did any speed test since my RTX 4080 is clearly faster.

Cu Selur

Ok, I'd like to understand how feasible (video) inferences are WITHOUT a discrete video card...

EDIT
I forgot to mention that I've collected some "AI" (= deep learning/neural nets) video tools in this GitHub page.
Suggestions - and direct contribution too - are welcome.

[Selur]

Did anyone manage to get BasicVSR++ working with R66, now R67? (both requiring Python 3.12)

[Selur]

(since a user asked me via pm)
For those interested, here's how I atm. would create a portable Vapoursynth 67 with pytorch&co:

• creat a new empty Vapoursynth-folder
• download 'Windows embeddable package (64-bit)' from https://www.python.org/downloads/release/python-3122/
• extract the Python download into the download into the 'Vapoursynth'-folder
• downloaded 'VapourSynth64-Portable-R67' from https://github.com/vapoursynth/vapoursynth/releases
• extract the Vapoursynth portable download into the 'Vapoursynth'-folder
• download get-pip.py from https://bootstrap.pypa.io/get-pip.py and save it into the 'Vapoursynth'-folder
• open a 'Windows Command Prompt'-window and navigate into the 'Vapoursynth'-folder
• install pip by calling :
  

  python get-pip.py

• open the python312._pth in a text addition and added the following to lines above anything else in that file and saved the file
  

  Scripts
  Lib\site-packages

• install wheel
  

  python -m pip install wheel/VapourSynth-67-cp312-cp312-win_amd64.whl

• installed pytorch using:
  

  python -m pip install torch===2.2.0+cu121 torchvision===0.17.1 -f https://download.pytorch.org/whl/torch_stable.html

• TensorRT
  • created a 'torch_dependencies/bin'-folder inside the Vapoursynth folder
  • download https://developer.download.nvidia.co....cuda-12.4.zip
  • extract (from the lib folder into the 'torch_dependencies/bin'-folder):
    • nvinfer.dll
    • nvinfer_builder_resource.dll
    • nvinfer_plugin.dll
    • nvinfervc__plugin.dll
    • nvonnxparser.dll
  • extract (from the python folder into the Vapoursynth folder: 'tensorrt-10.0.0b6-cp312-none-win_amd64.whl' and called:
    

    python -m pip install tensorrt-10.0.0b6-cp312-none-win_amd64.whl

  • download https://developer.download.nvidia.co...78_windows.exe
  • extract:
    • cuda_cudart\cudart\bin\cudart64_12.dll
    • cuda_nvrtc\nvrtc\bin\nvrtc64_120_0.dll
    • cuda_nvrtc\nvrtc\bin\nvrtc-builtins64_124.dll
    • libcublas\cublas\bin\cublas64_12.dll
    • libcublas\cublas\bin\cublasLt64_12.dll
    • libcublas\cublas\bin\nvblas64_12.dll
    • libcufft\cufft\bin\cufft64_11.dll
    • libcufft\cufft\bin\cufftw64_11.dll
    • libcurand\curand\bin\curand64_10.dll
    • libcusolver\cusolver\bin\cusolver64_11.dll
    • libcusolver\cusolver\bin\cusolverMg64_11.dll
    • libcusparse\cusparse\bin\cusparse64_12.dll
    • libnvjpeg\nvjpeg\bin\nvjpeg64_12.dll
    into the 'torch_dependencies/bin'-folder
• download https://developer.download.nvidia.co....0_windows.exe
• extract:
  • cudnn_cuda12.4\libcudnn\bin\12.4\cudnn64_9.dll
  • cudnn_cuda12.4\libcudnn\bin\12.4\cudnn_adv64_9.dll
  • cudnn_cuda12.4\libcudnn\bin\12.4\cudnn_cnn64_9.dll
  • cudnn_cuda12.4\libcudnn\bin\12.4\cudnn_ops64_9.dll
• call

  set PATH=<PATH TO torch_dependencies>\bin;%PATH%

  i.e.:

  set PATH=F:\Vapoursynth\torch_dependencies\bin;%PATH%;

• install VSGAN (https://github.com/rlaPHOENiX/VSGAN)
  

  python -m pip install -U VSGAN

• installed DPIR (https://github.com/HolyWu/vs-dpir)
  • python -m pip install -U vsdpir
    python -m pip install packaging
    python -m pip install onnxruntime-gpu
    python -m vsdpir[/code]
• install Real-ESGRAN (https://github.com/HolyWu/vs-realesrgan)
  

  python -m pip install -U vsrealesrgan
  python -m vsrealesrgan

• install vs-rife (https://github.com/HolyWu/vs-rife)
  

  python -m pip install --upgrade vsrife
  python -m vsrife

• install GMFSS (https://github.com/HolyWu/vs-gmfss_fortuna)
  

  python -m pip install -U vsgmfss-fortuna

• install SwinIR (https://github.com/HolyWu/vs-swinir)
  

  python -m pip install -U vsswinir
  python -m vsswinir

• install HINet (https://github.com/HolyWu/vs-hinet)
  

  python -m pip install --upgrade vshinet
  python -m vshinet

• install AnimeSR (https://github.com/HolyWu/vs-animesr)
  

  python -m pip install -U vsanimesr

• install FeMaSR (https://github.com/HolyWu/vs-femasr)
  

  python -m pip install -U vsfemasr
  python -m vsfemasr

• install SCUNet (https://github.com/HolyWu/vs-scunet)
  

  python -m pip install -U vsscunet
  python -m vsscunet

• install CodeFormer (https://github.com/HolyWu/vs-codeformer)
  

  python -m pip install -U vscodeformer
  python -m vscodeformer

• install GRLIR (https://github.com/HolyWu/vs-grlir)
  

  python -m pip install antlr4-python3-runtime
  python -m pip install wheel
  python -m pip install -U vsgrlir
  python -m vsgrlir

• install BasicVSR++ (https://github.com/HolyWu/vs-basicvsrpp)
  ????

Sadly I have no clue how to BasicVSR++ (https://github.com/HolyWu/vs-basicvsrpp) installed, so far

set CUDA_HOME=F:\Vapoursynth\torch_dependencies\bin;
python -m pip install -U openmim==0.3.7

anything I tried so far after this fails :/
Seems like the problem is that there seems to be no mmcv with cp312 wheel for Windows.

Hope this is useful to anyone and if anyone finds a way to add BasicVSR++

Cu Selur

[Cary Knoop]

Thanks for that detailed description!

Life could be a lot easier if VapourSynth simply adhered to common standards.

More than a decade of VapourSynth and a simple "pip install VapourSynth" from a conda environment is still not supported.

[Selur]

Only 'fault' with Vapousynth is that it abandoned Python 3.11 and now requires Python 3.12, which other projects do not support atm.

[PatchWorKs]

Any (Adobe-backended) VideoGigaGAN test ?

Demo: https://videogigagan.github.io/assets/videos/demo.mp4

Project: https://videogigagan.github.io/

Git: https://github.com/danaigc/videoGigaGanHub

[Selur]

https://github.com/danaigc/videoGigaGanHub seems to lack the used models, so no way to reproduce the presented results (or test with other content).

[PatchWorKs]

...well, it seemed a bit strange that Adobe-backeded researchers pretended to be open (using Microsoft's GitHub, moreover)...

[PatchWorKs]

Fast-SRGAN (The goal of this repository is to enable real time super resolution for upsampling low resolution videos. [] For speed, the upsampling is done through pixel shuffle.) anyone ?

[Selur]

Isn't that like 5 years old,... with https://github.com/HolyWu/vs-realesrgan I get way faster speeds on 180x320 content, but then my hardware is also way faster than a m1.

[Selur]

Did anyone try ccrestoration ?

[ReinerSchweinlin]

No - but the amount of usable models look very interesting.

It seems the open models site offers the models in pytorch, ONNX and NCNN Format - Isn´t hybrid already capable of using some of these ?

[Selur]

Yes, Hybrid can use those that VSGAN and VSMLRT support, but trying somthing new seems worth a look especially since some of the mentioned architectures https://github.com/TensoRaws/ccresto...n/type/arch.py are not supported by VSGAN and VSMLRT afaik.

[poisondeathray]

vs_liff - interesting resizing project . Works ok , but looks like only the baseline model so far

https://github.com/pifroggi/vs_liif

https://yinboc.github.io/liif/

[Selur]

Looked at it a while ago, but couldn't find an example where I liked the output

[ReinerSchweinlin]

Originally Posted by Selur

Yes, Hybrid can use those that VSGAN and VSMLRT support, but trying somthing new seems worth a look especially since some of the mentioned architectures https://github.com/TensoRaws/ccresto...n/type/arch.py are not supported by VSGAN and VSMLRT afaik.

Tempting

[ReinerSchweinlin]

Originally Posted by Selur

Looked at it a while ago, but couldn't find an example where I liked the output

For some Line Art it produces not so bad results and IMHO it could be usefull as a generic upscaler - while most GAN based algos are much better in cases where they are trained on the spcific footage and "recognize" the content and add plausible details - in cases where the don´t recognize anything, the often fall apart (zombie faces) - here the liif apporach could work better.

[asarian]

This is sad. I was hoping to use vs-basicvsr, to replace Oyster (as it no longer works); and now this one is gone altogether. And I really need a filter to get rid of those ugly compression artifacts.

[Selur]

vs-basicvsr was replaced with BasicVSR++ quite a while ago,...
(answered in your Oyster thread how to fix the Oyster.py to work with newer Vapoursynth versions)