Search Results
12/19/2024, 12:41:29 PM
>>8370503
>That is probably not feasible, as it would take a very long time to extract the information from all the frames, aswell as the amount of storage. And if you use a small portion, it wouldn't really work, since hentai has so many identical frames.
You can't honestly say this when Nvidia DLSS exists and does upscaling for a video game which has tight requirements on real time input and timing on showing a frame on a screen. Video by comparison most of the time runs slower than half what a "well-running" video game is required to do at 60 FPS. There is so much that can be done given double the time of doing a video game tier upscale that even without the specialized inputs, it's not like you couldn't at least get close enough to that quality if all the pieces were there. It's an infrastructure problem really and I will continue to maintain it is the case.
>>8370509
>Realtime is honestly nothing important IMO, as long as we don't have to wait an eternity for an upscale. Like my upscales which uses Ani4K+AnimeVideoV3 which run at 0.404x of realtime (assuming 23.976 fps).
Define "eternity". We already have consternation on upscales of a single image taking double digit of seconds time to do for a single frame in prior posts. It's quite clear that it does matter. If we're talking about using state of the art, it actually takes a full minute to do an upscale with an academic model of a single frame. We're talking about 6.951×10^-4 speed or 3 whole decimal digits of magnitude difference.
Refer back to how I qualify efficiency as not only being of model size but of runtime and operations used. But I get you. Even though the original ESRGAN paper was written 6+ years ago and we've gotten much better models, it is worth asking if the juice is worth the squeeze. The state of the art model has the following numbers.
>1
>Hi-IR-L
>SSIM 0.9366
>PSNR 33.13
Is the .02 SSIM and ~1.5 PSNR gain on paper, not judged qualitatively, worth that?
(2/2)
>That is probably not feasible, as it would take a very long time to extract the information from all the frames, aswell as the amount of storage. And if you use a small portion, it wouldn't really work, since hentai has so many identical frames.
You can't honestly say this when Nvidia DLSS exists and does upscaling for a video game which has tight requirements on real time input and timing on showing a frame on a screen. Video by comparison most of the time runs slower than half what a "well-running" video game is required to do at 60 FPS. There is so much that can be done given double the time of doing a video game tier upscale that even without the specialized inputs, it's not like you couldn't at least get close enough to that quality if all the pieces were there. It's an infrastructure problem really and I will continue to maintain it is the case.
>>8370509
>Realtime is honestly nothing important IMO, as long as we don't have to wait an eternity for an upscale. Like my upscales which uses Ani4K+AnimeVideoV3 which run at 0.404x of realtime (assuming 23.976 fps).
Define "eternity". We already have consternation on upscales of a single image taking double digit of seconds time to do for a single frame in prior posts. It's quite clear that it does matter. If we're talking about using state of the art, it actually takes a full minute to do an upscale with an academic model of a single frame. We're talking about 6.951×10^-4 speed or 3 whole decimal digits of magnitude difference.
Refer back to how I qualify efficiency as not only being of model size but of runtime and operations used. But I get you. Even though the original ESRGAN paper was written 6+ years ago and we've gotten much better models, it is worth asking if the juice is worth the squeeze. The state of the art model has the following numbers.
>1
>Hi-IR-L
>SSIM 0.9366
>PSNR 33.13
Is the .02 SSIM and ~1.5 PSNR gain on paper, not judged qualitatively, worth that?
(2/2)
Page 1