Using NV's Tensorcores to accelerate Waifu2x and upscale games in realtime. #427
fsjgsklgksdh
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Not sure whether this has already been done and whether this will be useful to anyone.
I wrote some code to utilize Nvidia's Tensorcores to accelerate waifu2x up-scaling and tried to tune it in order to achieve performance as high as possible. The whole point of this was to upscale some low resolution anime games in real-time and play them at higher resolution. I guess this can be used to watch anime videos as well.
With an Nvidia RTX 3070ti and while up-scaling from 800x600 to 1600x1200 I was able to achieve a performance of around 45 frames per second. This means that I could even play some old waifu2x up-scaled FPS shooters. (Although I did it just for the sake of trying and I do not recommend repeating that because they are not in anime style. But this might be well suitable for games like visual novels.) Some sacrifices I had to make to achieve this performance was the disabling of the de-noising and switching to 16 bit floating point accuracy. But the overall resulting quality was still good. I could not see any degradation of quality with my eyes, but I also have to mention that my input did not require any de-noising.
Does this sound interesting or is the reaction rather "Meh"? I'm asking because I could publish my code, but if there is no interest then there is no point for me to waste my time and taking the effort.
I wrote this message with a disposable account, so do not write me private messages. If there is enough interest I will take some time to publish the results properly.
Beta Was this translation helpful? Give feedback.
All reactions