21.02.2023, 20:03
Hello Selur,
I performed some tests on the 2 most interesting AI resizer filters available in Hybrid, in the case the results can be useful to other people looking for good resizers.
The 2 AI resizers that I tested are FeMaSR and realesr-general (RealESRGAN model suitable for videos). The first is known to be able to produce very sharp videos with enhanced details. The second is able to produce very clean images with less details respect to FeMaSR.
A comparison of the 2 models is available at the following link.
https://imgsli.com/MTU3MDU0
as it is possible both the models have problems: FeMaSR video is too sharp and RealESRGAN is too smooth.
Thanks to the weighed resizer it is possible to mitigate the sharpness of FeMaSR, but how to improve the sharpness of RealESRGAN ?
I tried the following approach: I apply the CASGLSL before the resize and then apply a CAS (with CUDA) after the resize.
In the following link there is a comparison of the 2 models, where I applied a weighed resize of 65% on both (the model output is reduced by 35%).
https://imgsli.com/MTU3MDU5
The defects of the 2 models has been reduced significantly (can be reduced more by changing the weight), more interesting it was possible to add the sharpness to RealESRGAN model. Even more interesting the encoding speed of realesr-general is 3x faster than FeMaSR.
I don't pretend that these results can be considered applicable to all the videos, unfortunately the behavior of these models can vary significantly from one video to another, but it could provide some insight on the results that one could expect to obtain using these models.
Dan
I performed some tests on the 2 most interesting AI resizer filters available in Hybrid, in the case the results can be useful to other people looking for good resizers.
The 2 AI resizers that I tested are FeMaSR and realesr-general (RealESRGAN model suitable for videos). The first is known to be able to produce very sharp videos with enhanced details. The second is able to produce very clean images with less details respect to FeMaSR.
A comparison of the 2 models is available at the following link.
https://imgsli.com/MTU3MDU0
as it is possible both the models have problems: FeMaSR video is too sharp and RealESRGAN is too smooth.
Thanks to the weighed resizer it is possible to mitigate the sharpness of FeMaSR, but how to improve the sharpness of RealESRGAN ?
I tried the following approach: I apply the CASGLSL before the resize and then apply a CAS (with CUDA) after the resize.
In the following link there is a comparison of the 2 models, where I applied a weighed resize of 65% on both (the model output is reduced by 35%).
https://imgsli.com/MTU3MDU5
The defects of the 2 models has been reduced significantly (can be reduced more by changing the weight), more interesting it was possible to add the sharpness to RealESRGAN model. Even more interesting the encoding speed of realesr-general is 3x faster than FeMaSR.
I don't pretend that these results can be considered applicable to all the videos, unfortunately the behavior of these models can vary significantly from one video to another, but it could provide some insight on the results that one could expect to obtain using these models.
Dan
