227 Pages with 35 video tutorials.
Why it matters:
Generated results cannot be evaluated only by looking at them, but automatic metrics are also imperfect.
Topics:
FID
CLIP score
FVD
LPIPS
Human preference models
Face identity similarity
Lip-sync metrics
Why metrics often disagree with human judgment
Suggested materials:
FID paper
CLIPScore
Fréchet Video Distance
LPIPS
PickScore / ImageReward / HPSv2
SyncNet / Wav2Lip evaluation