Announcement
SemanTok models are claimed to beat larger VideoFlexTok models on semantics and fidelity
A contributor to a Stability AI summer project describes SemanTok as using predictable semantic tokens for autoregressive video generation.
TLDR
A contributor to a summer project at Stability AI introduced SemanTok and claims its 49M autoregressive model beats VideoFlexTok models up to 47 times its size on semantics. They also claim a 201M model beats a 3.4-times-larger VideoFlexTok model on fidelity.
Combined views
1.1K
2 Sources, first seen ago
