I will have to hardcore test this with my test images.

Also how does Lanczos filtering work again? I need this code a little more commented.

idk how to explain lanczos, my brain is fried. https://en.wikipedia.org/wiki/Lanczos_resampling

Hope that comment explains it

Can we name some variables better, like what is a meant to be?

Also is lanczos_window basically square_block_size like in the code below? (I'm just trying to understand exactly how this works, I'll read more into that wiki later™)

Alright starting to understand this...

EDIT: I'm commenting more of the code now.

I don't know if it's gotten slower or what but image-rs down-sampling is faster with every other image except purposely large images like uwu_bomb.png. Lanczos is naturally more expensive so I expect it to be slower than the previews algorithm but now I got to compare the two versions and see if I can somehow improve it.

Imagine this on an AMD Threadripper 💀 you'll be able to downsample 8k images A LOT faster.

BTW I do have a potential solution I could try to speed this up, i just don't know I'll work until I actually implement it and test it A LOT as we might be talking about very small milliseconds of improvement.

performance issues fixed! It's a lot faster than the image-rs backend now with all images I've tested so far.

During my tests I found the roseate backend having issues with some images

going back versions this issue has existed since I wrote the first iteration though so I can still close this pr, thanks!