Physics report card: what each video model gets WRONG, so you can compose around it
Spent a month cataloguing the physics failures per model so I can pick the right tool and dodge its weakness. The big ones:
Hair/cloth: Everyone struggles. Kling best, but long hair in motion still occasionally phases through a shoulder. Compose with tied-up hair or minimal cloth movement for safety.
Water: Seedance best (that film look extends to beautiful water), Runway worst (water often moves like gelatin, surface tension wrong). Hailuo's water is okay if it's not the subject.
Fire/smoke: Kling and Hailuo good, Runway's smoke can be weirdly solid/clumpy. Fire is generally safer than smoke across all of them.
Weight/gravity in action: Nobody's great. A jump, a fall, a thrown object — the arc is usually subtly wrong (too floaty, wrong acceleration). AVOID dynamic action if you want believability; this sub is about the HELD moment anyway, which conveniently dodges the whole problem.
Reflections (water/mirror/glass): All weak, same as the still-image models. The reflection is vibes. Keep reflective surfaces non-hero or expect to composite.
The meta-lesson: compose TO the model's strengths. A locked, minimal-motion, atmospheric mood piece dodges 80% of these. The failures all live in dynamic action and complex physics, which the single-shot mood piece naturally avoids.