Search Results
12/18/2024, 11:03:52 AM
>>8367767
I haven't done much A to B testing for accuracy as far as hue shadeifferences but yeah, I agree, RPLKSR then is not ready yet as you say with the issues with the background. But as I said, no real great models are really ready and for my needs, it works alright even if the architecture shows promise and even the author went back to ESRGAN with AnimeSharpV3 which superseded it. I disagree with the magnitude difference, RPLKSR was 10.3 seconds vs almost 1.9 seconds for me than with RealESRGAN but neither are really running realtime here and I'm probably running a faster GPU than you are.
I disagree strongly with your assessment that DAT merely trades blows with the compact model. The reconstruction is clearly better, if you for example look at one of your comparisons with Serial Lain Experiments, for example. The loops around the patch has clear artifacting with inconsistent shaded colors and patches of white even inside those patch holes/loops on the ESRGAN model while on the DAT side, there is none of it and you'll see similar issues in high frequency areas that ESRGAN has a harder time upscaling that gets a much cleaner result with DAT. Moreover, the reconstruction on the drawn lineart is aesthetically better in my opinion than with ESRGAN's reconstruction. I really dislike the way ESRGAN reconstructs lines, it's usually looks like a crude approach to vectorizing the image. This is why I prefered the RPLKSR model too where the line reconstruction actually looks like it could be drawn while it doesn't look great with Real-ESRGAN.
Don't get me wrong, Real ESRGAN and its finetunes are still pretty great, but this is where state of the art upscaling with transformer models have started to yield clear better results if you care about quality and can afford that extra time to get better details for single images. But for videos, that additional time for using DAT in a potential model does add up per frame so it's probably not that worth it to do just yet.
I haven't done much A to B testing for accuracy as far as hue shadeifferences but yeah, I agree, RPLKSR then is not ready yet as you say with the issues with the background. But as I said, no real great models are really ready and for my needs, it works alright even if the architecture shows promise and even the author went back to ESRGAN with AnimeSharpV3 which superseded it. I disagree with the magnitude difference, RPLKSR was 10.3 seconds vs almost 1.9 seconds for me than with RealESRGAN but neither are really running realtime here and I'm probably running a faster GPU than you are.
I disagree strongly with your assessment that DAT merely trades blows with the compact model. The reconstruction is clearly better, if you for example look at one of your comparisons with Serial Lain Experiments, for example. The loops around the patch has clear artifacting with inconsistent shaded colors and patches of white even inside those patch holes/loops on the ESRGAN model while on the DAT side, there is none of it and you'll see similar issues in high frequency areas that ESRGAN has a harder time upscaling that gets a much cleaner result with DAT. Moreover, the reconstruction on the drawn lineart is aesthetically better in my opinion than with ESRGAN's reconstruction. I really dislike the way ESRGAN reconstructs lines, it's usually looks like a crude approach to vectorizing the image. This is why I prefered the RPLKSR model too where the line reconstruction actually looks like it could be drawn while it doesn't look great with Real-ESRGAN.
Don't get me wrong, Real ESRGAN and its finetunes are still pretty great, but this is where state of the art upscaling with transformer models have started to yield clear better results if you care about quality and can afford that extra time to get better details for single images. But for videos, that additional time for using DAT in a potential model does add up per frame so it's probably not that worth it to do just yet.
Page 1