usage: UniversalTorchUpscaler.py [-h] [-i INPUT] [-o OUTPUT] [-t TILESIZE]
[-b BACKEND] [-m MODELPATH] -n MODELNAME [-c]
[-f FORMAT] [--half] [--bfloat16] [-e EXPORT]
Upscale any image, with most torch models, using spandrel.
options:
-h, --help show this help message and exit
-i INPUT, --input INPUT
input image path (jpg/png/webp) or directory
-o OUTPUT, --output OUTPUT
output image path (jpg/png/webp) or directory
-t TILESIZE, --tilesize TILESIZE
tile size (default=0)
-l OVERLAP, --overlap OVERLAP
overlap size on tiled rendering (default=10)
-b BACKEND, --backend BACKEND
backend used to upscale image. (pytorch/ncnn, default=pytorch)
-m MODELPATH, --modelPath MODELPATH
folder path to the pre-trained models. default=models
-n MODELNAME, --modelName MODELNAME
model name (include extension)
-c, --cpu use only CPU for upscaling, instead of cuda. default=auto
-f FORMAT, --format FORMAT
output image format (jpg/png/webp, auto=same as input, default=auto)
--half half precision, only works with NVIDIA RTX 20 series and above.
--bfloat16 like half precision, but more intesive. This can be used with a wider range of models than half.
-e EXPORT, --export EXPORT
Export PyTorch models to ONNX and NCNN. Options: (onnx/ncnn)