Differential Mean Opinion Score (DMOS) measures how much a clip degraded relative to its pristine reference, not how good it looks on its own. It exists to cancel a confound in plain MOS, where a viewer's rating mixes the impairment with how much they liked the content and with their personal harshness; DMOS always measures each viewer against a reference they themselves rated. There are two conventions, and confusing them is a real error source. In ITU-T P.910's ACR-HR method the per-viewer differential is DV = V(PVS) minus V(REF) plus 5, averaged into a DMOS where higher means better. In academic databases like LIVE, a per-subject difference is Z-scored and rescaled, often to 0-100, where higher usually means worse, the opposite direction. Same three letters, inverted meaning, so always state the convention and direction. VMAF was trained on database-style DMOS, which is why a VMAF point predicting a DMOS point depends on knowing which one.
