Deoldify Vapoursynth filter

[Dan64]

Wow! vs-realesrgan v5.1.0 is much faster.
On my PC with RTX3060 the speed increased from 12.95 fps to 30.38 fps (num_streams=3, num_batches=1) an increase of about 2.3x.

With num_batches=2 (num_streams must be =1 to improve the num_batches performance) the speed increased to 34.15 an increase of about 2.6x.
Probably due to my poor GPU I don't have observed any advantege in using num_batches > 2.

In any case it is an incredible result, with this speed it is possible to perform the resize with RealESRGAN in real-time.
I remember the times where the resizing speed of RealESRGAN was about 3/4 fps.

Dan


P.S.
Using the version 5.1.0 is displayed the following message:

Unable to import quantization op. Please install modelopt library (https://github.com/NVIDIA/TensorRT-Model...stallation) to add support for compiling quantized models

[Selur]

Using the version 5.1.0 is displayed the following message:

Unable to import quantization op. Please install modelopt library (https://github.com/NVIDIA/TensorRT-Model...stallation) to add support for compiling quantized models

=> you would have to ask HolyWu about that. (got everything installed that was listed in README)

Cu Selur

[Dan64]

Using the version 5.1.0 is displayed the following message:

Unable to import quantization op. Please install modelopt library (https://github.com/NVIDIA/TensorRT-Model...stallation) to add support for compiling quantized models

=> you would have to ask HolyWu about that. (got everything installed that was listed in README)

issue open: https://github.com/HolyWu/vs-realesrgan/issues/31


Dan

EDIT:
HolyWu answer: None of ModelOpt's functionality is used so you can safely ignore that message.