How well does Google Gemini Omni Flash follow time-specific prompts compared to Veo 3.1?

Google Omni Flash follows time-specific prompts more accurately than Veo 3.1 — across 30+ tested outputs, time code accuracy when prompting for specific time codes was consistently sharper in Omni. Simple timed action staging lands reliably; complex physics tied to a time code runs roughly 50/50, so plan rerolls for physics-critical beats.

Omni Flash beats Veo 3.1 on time code adherence, based on 30+ test outputs generated to evaluate the model — when you prompt "at 2 seconds the door opens, at 5 seconds she turns," Omni hits those marks more often than Veo 3.1 does. Write time-specific prompts as explicit second markers or bracketed second ranges rather than vague sequencing words like "then" or "after that" — the model resolves numbers better than relative ordering.

Work inside the clip-length windows. Omni generates at 4, 6, 8, or 10 seconds only, so anchor your time codes to one of those durations — a beat prompted at 9 seconds inside an 8-second generation simply gets dropped or compressed.

Split your expectations by prompt complexity. Simple staged actions at a time code perform consistently. Complex physics prompted at a specific moment — collisions, fluid, chained cause-and-effect — is roughly 50/50 in accuracy, but the correct outputs are exceptional, so budget multiple generations for those shots rather than treating the first miss as a model failure.

Two timing-adjacent caveats. First, camera angle changes tied to a time code are hit-or-miss in Omni, and a failed angle change tends to distort the scene geography rather than just the angle — keep angle switches and timed action in separate prompts where you can. Second, neither model will execute a timed beat involving real contact-based actions or anything resembling violence; that limitation is consistent across all Veo models, so no time code phrasing gets around it.

One workflow note: extend currently works only on clips generated in Veo 3.1, not Omni-generated content — if your timed sequence needs to grow past the 10-second cap, that pulls Veo 3.1 back into consideration despite the weaker time code accuracy. Both models run inside invideo, so you can send the same time-coded prompt to each and compare outputs directly, and the invideo agent routes each shot to whichever model fits the beat.

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, on Omni's complex physics prompt handling