Yeah, those two demos fell flat for me. The model performing badly on inputs far outside the training data is fine, but those two videos belong in the outtakes

Yeah, those two demos fell flat for me. The model performing badly on inputs far outside the training data is fine, but those two videos belong in the outtakes section or maybe a limitations section, not next to text lauding the "out-of-distribution generalization capabilities". The videos show the opposite of what's claimed.