1. It should follow the usual convention for rounding, .5 is rounded upwards. (It appears to be so ... but don't nail me on that it does so in each possible case. I'm not the coder of MaskTools.)
2. A pixel can't have a negative value, only 0-255. In order to handle "negative" differences, the range -127..0..127 is offsetted to 0..128..255.
3. You tell, I don't know. The exact operation of ML3Dex isn't fully clear to me (have been a bit lazy when looking through that pdf) ... however in practice, it doesn't impress me too much. The temporal artefacts in motion areas (resp. areas with erroneous motion compensation) are pretty much the same as those of plain temporal median, so there's no benefit in that respect. In areas without motion (resp. in areas with correct motion compensation), it does remove a bit more signal spikes, no matter whether it's noise or detail.
Thus, for me ML3Dex isn't a interesting option. And ML3DexGPU isn't because I don't have GPU acceleration.
If it works sufficient for you, then you answered your own question.
