Why does Gemini Omni Flash get physics prompts wrong in AI-generated videos?

Gemini Omni Flash fails physics prompts for three main reasons: complex physics scenarios resolve at roughly 50/50 accuracy in testing across 30+ outputs, safety filters block or soften any contact-based or violence-adjacent action, and failed camera-angle changes distort scene geography — breaking spatial logic even when object motion itself is correct.

Complex physics is a probabilistic weak spot, not a blanket one. Across 30+ tested outputs, simple physics adherence in Omni Flash was strong — objects fall, bounce, and interact plausibly in straightforward prompts. Complex physics prompts (multi-object interactions, chained cause-and-effect) landed at roughly 50/50 accuracy: sometimes correct, sometimes wrong, with no reliable way to predict which. When the model gets it right, the results are exceptional — which is why the same prompt can produce one convincing take and one broken one.

Safety filtering is the second failure mode, and it looks like a physics failure even though it isn't. Omni will not generate real contact-based actions or anything resembling violence — a consistent limitation across all Veo-family models. A physics prompt involving impact, collision, or force applied to a person often gets blocked or quietly softened, so the output shows objects stopping short, phasing past each other, or resolving without contact. If your physics prompt keeps failing, check whether the interaction reads as violence-adjacent and rephrase it around inanimate objects or indirect contact.

Spatial logic breaks compound the problem. Camera-angle prompting in Omni is hit-or-miss, and failed angle changes tend to distort scene geography rather than just the angle — so a shot can have correct object motion but a warped space around it, which reads as a physics error. Motion timing is also not fully stable: stop-motion-style outputs oscillated between 12 FPS and 8 FPS within clips, showing the model's temporal consistency isn't locked either.

Notably, the failure isn't prompt comprehension. Time code accuracy when prompting for specific time codes is sharper in Omni than in Veo 3.1 — the model understands what you asked and when; it's the physics simulation itself that misfires. The practical response to a 50/50 failure rate is volume and routing: generate multiple takes of any physics-heavy shot and keep the correct one, and inside invideo — where Veo, Kling, and Seedance 2.0 all run alongside Omni — the invideo agent can route a shot that keeps failing to a different model rather than forcing it through the same one.

Watch some of these to see what works for you:

this is kind of 50/50. It gets it right sometimes, it gets it wrong sometimes, but what it gets right, it gets it really right.

— invideo's creative team, after testing 30+ Omni Flash outputs