-
They outperform ravu, FSRCNNX and krigbilateral in several tests here. |
Beta Was this translation helpful? Give feedback.
Replies: 6 comments 17 replies
-
I have previously tested the effects of the ArtCNN shader, and it shows no significant difference in upscaling of HD content compared to FSRCNNX. It is only slightly stronger in upscaling of SD content than FSRCNNX (both effect poorly, many artifacts). Don't blindly trust the results of numerical tests, as the ravu ar shader that performed well in that test also performed poorly in upscaling SD content (some artifacts). So you can see that the upscaling shaders used as benchmarks there are not very recommended in my configuration. Of course, in terms of upscaling high-definition content, the credibility of that numerical test is still good |
Beta Was this translation helpful? Give feedback.
-
By content are you referring to anime or live action? Where in your opinion does nnedi3 fit in, similar to FSRCNNX? Good on hd and bad on sd if so why not remove it too? Also what are your thoughts on Cfl or the chroma variant of ArtCNN as a chroma replacement for krigbilateral? My point in asking these questions is to find the best combination of shaders for anime and live action for both sd and hd content without having a shader folder with 20+ shaders. |
Beta Was this translation helpful? Give feedback.
-
both, mainly animation.
The biggest feature of nnedi3 is that it hardly produces artifacts during upsampling and suppresses the original artifacts in the image, making its versatility irreplaceable. It is surprising that after so many years, there is still no better algorithm to replace it. The reason why FSRCNNX is retained is because the poor upscaling effect of SD content is almost a common problem among other shaders except for specific shaders, and its effects on HD content is indeed good.
The EASU_rgb variant of FSR often effect similarly to nnedi3 and has lower performance consumption, which is why it is retained. I also do not recommend using Anime4k for HD anime as it will disrupt the original details. But its effects on SD anime is significantly better than any other shader, with powerful anti aliasing and artifact removal capabilities. You may have noticed that the Anime4k usage I recommended in the configuration does not include its upscale algorithm. Lines 757 to 759 in 014c9db
I really don't care much about chroma upscaling. The human eye has low sensitivity to chroma information and it's not worth investing too much effort and performance in it. Krigbilateral is also unnecessary. The method of inferring chroma information through luma information using Cfl shaders may not work well when there is no strong correlation between luma and chroma. But once again, I don't care about chroma upscaling, so I won't test and compare them. The above is just a subjective judgment.
I won't recommend the best combination for anyone because everyone's preferences are different. Shaders with poor effects on SD content have no impact on people who only watch HD content, and for those who like image sharpening, they will not like shaders like nnedi3. So you can only rely on yourself to choose and use the shader scheme that suits you. |
Beta Was this translation helpful? Give feedback.
-
I turned it into a discussion. Considering that ArtCNN is still actively developing and FSRCNNX is no longer maintained, perhaps one day ArtCNN will be enough to replace FSRCNNX. PS: simplify ravu shaders in new commit: 2c91112 |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
After some recent testing, I found a better combination of universal shader configurations, see the commits: b9571c5, 3742d70, 4eaace6 And for animated scenes the best shader currently available is Ani4K. 91a1518 |
Beta Was this translation helpful? Give feedback.
I have previously tested the effects of the ArtCNN shader, and it shows no significant difference in upscaling of HD content compared to FSRCNNX. It is only slightly stronger in upscaling of SD content than FSRCNNX (both effect poorly, many artifacts).
There is no reason to add a similar and not significantly improved shader (both trained based on CNN models), Before this, I had already streamlined the shaders in the configuration for similar reasons: bb49059, 7cfec96
PS: The comparison here is the normal version of ArtCNN, whose DS variant is catastrophic in upscaling SD content.
Don't blindly trust the results of numerical tests, as the ravu ar shader that performed well in that test al…