16.12.2024, 22:22
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
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