Open fifi1ous opened 6 months ago
For training GS uses pic at 1600 x yyyy (only one size is 1600, yyyy is ratio dependent.) by default. You can choose any image size with -r parameter if you have enough VRAM.
And is there any way to find what rescaling algorithm GS uses? And what if the final yyyy isn't divisible by 16?
Some hint there ;) Do math from 1920x1020 to 1600x850 ratio is (1,882352941176471) in this example. 850 isn't divided by 16...
Yeah, I got this, but with the divisions I am talking about the rasterizer, which splits the screen into 16x16 tiles. Or am I getting it wrong and they dont split the picture screen, but the model on the screen.
It's not an issue but it is more a question about some parts of the process. One of the initial steps is to rescale the image into 1.6k x 1.6k, I assume that this step is because of the afterwards rasterizer process when they split the screen into 16x16 pixel tiles. I wonder why they chose the 16x16 pixel tiles? If someone chooses not to resize the images into different sizes than 1.6k x 1.6k, how does it affect the use of 16x16 pixel tiles? Also, what happens if the image dimensions aren't divisible by 16? Lastly, could you tell me what kind of resizing algorithm you use?