Time the spoken line before you generate the shot
A generated clip can look fine on its own and still be useless in an episode. The voice finishes, and the picture keeps going. Or the line gets cut off because the shot was a fixed 5 seconds and the sentence was longer.
A generated clip can look fine on its own and still be useless in an episode. The voice finishes, and the picture keeps going. Or the line gets cut off because the shot was a fixed 5 seconds and the sentence was longer.
We ran into this while building AlphaDrama, a web studio that turns a story into a vertical episode, or product photos into a short ad. The fix is ordinary, and it has to happen before the video model is called.
Measure the line first
- Write the line for that shot.
- Generate the voice and keep the audio file.
- Read the duration of that file.
- Ask for a shot of that length, not a round number you picked earlier.
If the line is 3.4 seconds, the picture should be about 3.4 seconds. Padding the shot "to be safe" is how you get a face staring at the camera after the sentence is over. Cutting the shot short is how you lose the last word.
Then keep the same person in the next shot
Timing does not fix a face that changes at shot 14. For that we make a character sheet first: front, three-quarter, and side. Every later shot reuses those images. The sheet is the reference. The prompt is not allowed to invent a new person.
Product ads have the same shape of bug. The bottle in shot 3 should be the bottle from the photo, including the words on the label. We lock that look from the uploaded photos before any shot is generated.
Captions are part of the shot
The export is 9:16, with captions on the picture. A lot of vertical video is watched with the sound off, so a line that exists only in the audio is a line a lot of people never get.
What we do not do
We do not generate the whole episode as one long clip and hope the person stays put. Each shot is its own generation, cut to the line, then stitched. When a shot is wrong, we redo that shot. We do not reroll the whole episode.
AlphaDrama is a web app. There is a free tier. Paid plans start at $12 a month.
- Story to a vertical episode: https://www.getalphadrama.com/ai-drama-tiktok-reels
- Product photos to a short ad: https://www.getalphadrama.com/ai-commerce-video-studio
If you cut multi-shot AI video, do you time the voice first, or do you trim the picture after the fact?
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.