We're also building our own model, for slides that actually look designed, and the hardest part was defining what "looks designed" even means. In the end we settled on a verbal definition of roughly "image-heavy" and "looks like a magazine spread," and then had no choice but to have humans label by feel.
That was manageable for us. Video sounds genuinely hard though: more senses involved, and I'd guess the variance is far larger.
That was manageable for us. Video sounds genuinely hard though: more senses involved, and I'd guess the variance is far larger.