libplacebo
libplacebo copied to clipboard
haasn
jinc scaler has errors.
jinc:
older version where utterly broken or i don't know how to reproduce such terrible results this test was not done by me:
https://slow.pics/s/xElVVHli
there are more scaler that show issue but i don't know which are entirely broken on such a test by design or just have bad weights like cosine, hamming, welch, sphinx and sinc.
even the normal lanczos shows issue.
the anti ringing seems to be not set up properly for many scaler too.
here is madVR jinc AR as an comparison:
if both are supposed to be the same scaler then the one in mpv needs to be fixed.
tested with 0.40.0-23-g38439dd53
MadVR's "Jinc" has its third lobe clipped at 3, and the AR algorithm is different.
clipped? i do not know that stuff in detail. it had a 4tab(that's a lob right?) version but that was terrible.
AR does not matter this much on this test image but yes it would be odd if they are the same. the issue are the blocks that's usually an issue with weight. the jinc2m(maybe has nothing todo with jincs just a name in that case) in mpcVR had the same problem but with changing the weights it got fixed. not saying it is a good scaler or anything but the scaler will ace this simple scaler test: