AI lip sync drifts on longer clips for two compounding reasons: models hold phoneme-to-mouth alignment only within a short generation window — testing across 30+ outputs on Google Omni Flash found 6–7 seconds is the consistent ceiling for a single speaker — and small per-frame timing errors accumulate rather than self-correct, so audio and mouth movement separate further the longer the clip runs.
The first mechanism is the per-generation sync window: a video model aligns mouth geometry to audio phonemes frame by frame, and that alignment holds reliably only for a few seconds before mouth shapes start lagging or leading the track. In structured testing of Google Omni Flash across 30+ outputs, "In our research we found that 6 or 7 seconds is kind of the ceiling of where the model is going to perform consistently well," per invideo's creative team — a model-level constraint, not something a better prompt fixes.
The second mechanism is drift accumulation. Long clips are produced by generating or stitching shorter internal segments, and each segment boundary introduces a tiny audio-to-mouth offset that never resets — errors compound instead of correcting, which is why a 30-second clip doesn't fail 5x worse than a 6-second one, it fails progressively worse toward the end. Community testing threads on AI video failure modes report the same pattern: degradation is predictable and cumulative, not random.
Multiple speakers accelerate the failure. "Multiple people talking in the same frame is still one of the things that we are seeing to be one of the greatest weaknesses of AI models in today's day and age" — this holds across every current model tested, not just Omni Flash, because the model must now hold two independent phoneme-to-face alignments in one generation. Compositing the left and right halves of the frame separately to fake a two-person conversation is a workaround, not a capability, and it shows at the seam.
What to do about it: generate dialogue as short single-speaker clips instead of one long take — Omni Flash's duration options of 4, 6, 8, and 10 seconds map almost exactly onto its 6–7 second reliable window, so treat 6–8 seconds as your per-take budget. Place cuts at natural speech pauses, and reset sync at each boundary by cutting to a reaction shot or B-roll rather than holding one continuous talking head. Note that scene extension currently only works on clips generated in Veo 3.1, not Omni-generated content, so you can't stretch an Omni dialogue clip past its window — plan the edit around short takes from the start. Inside invideo, where all current video models are available, you can generate each dialogue line as its own short clip and let the invideo agent assemble the sequence, which keeps every take inside its model's reliable sync window.
Watch some of these to see what works for you:
In our research we found that 6 or 7 seconds is kind of the ceiling of where the model is going to perform consistently well.
— invideo's creative team