What rights attach to talking-head and lip-sync checkpoints when the code is permissive but training data is restricted?

Wav2Lip's models were trained on LRS2 and its README prohibits commercial use, MuseTalk's README grants commercial use of weights while flagging test data as non-commercial, and SadTalker/Hallo2/EchoMimic state code licenses but not clear weights terms. Which checkpoints can be used for paid work and what evidence would settle it? Why this is still open: License statements are inconsistent across code, weights and datasets and several weights licenses are simply unstated.