I forgot to use the last torch version. I will update torch, but lowering the size of frame to be resized allows FeMaSR to resize the video.
For the same reason it is not possible to use FeMaSR_SRX4 if the ouput size is too big (1080p output size for me is not working).
Quote:For the same reason it is not possible to use FeMaSR_SRX4 if the ouput size is too big (1080p output size for me is not working).
Input size is more important. SD content uses ~8GB VRAM and 720p uses more than 16GB. for X2 with X4 it will require way more vram.
for some reason the femasr update didn't work properly, FeMaSR is still 'vsfemasr-1.0.0.dist-info'.
Managed to update it properly now, but packaging and uploading will take ~50min.
About the vram usage: you need to set 'trt_min_subgraph_size' to a higher value than the default 1. (I need to use 10.)
Code:
trt_min_subgraph_size: Minimum node size in a subgraph after splitting. Subgraphs with smaller size will
fall back to CUDA. Try trt_min_subgraph_size=5 if TensorRT engine cannot be built
due to insufficient VRAM, but the performance will degrade.
Cu Selur
Dear Selur.
Do you plan to include the FeMaSR update in the torchAddOn package
very interesting to try it.
Thank you.