mvtools does block-based estimation. Thus such a hack might not keep same strength/threshold for Y and UV, unless you have found a fine pel precision for each plane, which I should say is nearly impossible now, as pel must be 2^x. Anyway, not a bad idea for SD. For HD, one would rather do some hack on mvtools to make it support YV24 natively instead of using such a large block searching on luma.
Another useful idea is quite simple: do a multi-level search. Use a downsized clip as first-level, refine the vector with a larger clip as second-level, again refine the vector with a even large clip as third-level, recursively.... With such refining you can get exact same pel precision for Y and UV even you simply use luma to build the first-level and refines it on chroma, thought such a method is more reputed to be a stable vector builder for extremely grainy sources with a block-based motion estimation.
mawen1250 写了：但是在对PC Range的YCgCo进行bit depth转换时用的应该是TV Range的转换方法——bit shift，所以如果对其使用Down10、U10或U16时不应该指定tvRange=false。
Don't know where you've got this. Never documented. Any tool or any one working in this way does not necessarily mean it is correct, even if it seems to be sane.
feisty2 写了：第二点就是我还希望使用一下g_mask来逐渐减弱对暗处的降噪强度 但是taro大写的masktools都是8bit下的用法 如果直接按照8bit的用法来使用会有问题吗?
I don't think he limits his tutorial only in masktools. In that thread one of the key thoughts is merging two clips using the given value of the mask clip to control weight/alpha. There is no bit depth limitation in this thought.
Generally, YCgCo is not something good for filtering. It is known as a meaningful try rather for compression than for processing. Again as you've already made up your mind to do some lossy filtering, there is little meaning to use YCgCo just for some placebos such as the concept of lossless
. And if you're really addicted to the benefits of high depth or even floating-point filtering, or some terms referring to, uh, lossless
, give VapourSynth a shot, seriously.