There's going to be some bias in the photos and video that are available as training data. People are not normally going to keep shots where a person who's the focus has their face heavily in shadow.
The video shown had no monsters, magic, or similar. I do wonder how you can handle that sort of thing given that there can't be any real images to use for training.
The more notable artifact to me was the hat shadow vanishing: https://www.youtube.com/watch?v=4ZlwTtgbgVA&t=325s
There's going to be some bias in the photos and video that are available as training data. People are not normally going to keep shots where a person who's the focus has their face heavily in shadow.
The video shown had no monsters, magic, or similar. I do wonder how you can handle that sort of thing given that there can't be any real images to use for training.