Treat a spoken sentence as one event even when the picture changes. If both shot descriptions introduce the complete sentence as a new action, the generated result may restart it. Preserve one approved wording, identify where the camera cuts during it and distinguish an uninterrupted continuation from an intentional interruption.
Give the sentence its own identity
Write the whole utterance once in an editorial ledger with its speaker and intended ending. Below that, record which words belong before and after the picture cut. The ledger is a review aid, not a second set of lines to be spoken. Avoid repeating the full sentence in a summary, a soundscape field and each shot description.
The H3 base writing guide documents a continuation marker, <scenetrans>, for a line crossing a cut, together with explicit audio continuity. It separately identifies <cutoff> for truncation at the video's end. These are internal H3 prompt conventions; they are not editing commands, caption tags or universally supported website input syntax.
Separate camera boundaries from voice boundaries
Imagine an invented line: “Leave the blue folder beside the lamp.” The camera starts on the speaker and cuts to the desk during the sentence. The second view shows the folder but does not introduce a new speaker or repeat the instruction.
| Boundary | Required audio relationship |
|---|---|
| Speaker close-up to desk insert | Same utterance continues |
| Another person interrupts | Deliberate speech stop or overlap |
| Video ends before the sentence resolves | Revise duration or specify the approved truncation |
Before using a supported H3 rewrite pipeline, check the complete structure: integrated multimodal description, overall soundscape and non-diegetic music for base modes. Keep the approved dialogue in its intended dialogue representation, and use the provider's continuation rule at the connection. An explanatory fragment alone is not a complete runnable prompt.
Listen for three different failures
First, check duplicated words: the new shot may repeat the opening phrase or the boundary syllable. Second, check omissions: removing the repeat manually may delete a necessary word. Third, check voice continuity: the sentence can preserve every word yet change vocal identity, pace or room character at the cut.
Review the line with the picture hidden before judging the visual transition. Then restore the image and verify that the listener or insert shot does not begin mouthing an off-screen continuation. This is a proposed inspection method; no generation result or exact timing reliability is claimed here.
In an Imagild shot plan, state that the approved sentence continues through the camera change and keep the speaker unchanged. Use only input controls actually provided by the site. If generated clips cannot preserve that relationship, a supported external edit can retain a single approved audio performance across the cut. Judge the finished utterance, not merely the presence of a special marker in the brief.