Why do timestamped prompts cause dialogue hallucinations in AI video generation?

Timestamped dialogue prompts hallucinate because the model treats each timestamp as a time window it must fill with speech. When the scripted line ends before the window does, the model invents plausible filler — whispered half-lines, murmurs — rather than holding silence. Removing timestamps and giving the line plus scene context lets the model set its own pacing and eliminates the invented dialogue.

A timestamp in a dialogue prompt is a duration instruction, not a sync instruction. When you write "0:03–0:08 — 'We need to leave now'" the model reads a five-second speech obligation. If the line naturally takes two and a half seconds, the model still has dead time inside the window it was told contains dialogue — and it resolves that contradiction by generating more dialogue: whispered filler, invented reactions, murmured lines you never wrote. Documented Seedance 2.0 productions hit exactly this: timestamped dialogue prompts produced hallucinated whispered filler and wasted generation credits, while the same lines prompted without timestamps came back clean.

The fixed-length generation unit amplifies the problem. Video models generate in set-duration chunks — one documented production worked in 15-second clips — so a timestamp map across a full chunk almost always mismatches natural speech rhythm somewhere, and every mismatch is a gap the model fills with invented speech. Regeneration is the real cost: documented productions averaged 3 generations per usable shot before quality issues like this, so a prompting pattern that reliably injects wrong dialogue multiplies spend.

The fix is to prompt for context, not for a schedule. Give the model the exact line of dialogue, who says it, and the emotional register of the scene — then let the model determine pacing itself. One documented workflow states it directly: give it context, don't dictate second by second. For multi-line exchanges, split each character's lines into separate single-line clips; you get clean takes per line and full pacing control in the edit. Precise timing belongs in post anyway — in one documented production only about 5 seconds of each 15-second clip survived the cut, so trimming to your intended rhythm on the timeline is cheaper and more reliable than encoding rhythm into the prompt.

If you're generating through the invideo agent, run it in Always Ask mode so you review the assembled prompt before any credits are spent — confirm the dialogue block carries the line and scene context with no timestamp map before approving the generation.

Give it context, do not line by line, second by second, try to tell seedance what to say because you're not going to get the best results.

— an AI filmmaker documenting Seedance 2.0 dialogue prompting workflows