How to fix a new episode that drifted from your earlier ones
Point Vizard Agent at the earlier episode that got it right. It crops the title and caption areas out of that video and measures them, rebuilds the new episode's type to those measurements, compares the two side by side, and re-clones the narrator from a cleaner sample when the voice is the thing that drifted.
What do you type into Vizard Agent?
Name the episode instead of describing the style. "The font was larger, slightly blue and more central" is useful, but "make it like the opening of that other video" is precise, because the other video is a measurable object and your memory of it isn't. You don't need to know the font, size or colour: Vizard Agent measures those off the reference.
Prompt
What does Vizard Agent actually do?
A series drifts one small decision at a time: a slightly smaller title, a slightly different blue, a voice a shade flatter. Nobody notices in isolation, and everybody notices when two episodes play back to back. So almost every step is a measurement taken off the earlier videos. In the session behind this article it:
- Cropped the title and caption areas out of two earlier episodes of a vertical food-reel series, rather than relying on one, and inspected the typography close up.
- Measured the title wash timing and colour and the caption and title geometry, and checked how an earlier episode framed the same kind of shot.
- Matched the reference title's font size, identified the caption font, and compared the new title and the colour grades against the earlier episodes directly.
- Cloned the narrator from the earlier episodes' audio and built a voice comparison file to listen to the candidates.
- Re-cloned from a cleaner sample when the voice sounded synthetic, then verified the title and the voice on the finished render.
When does this not work well?
Matching needs something to match against. If the earlier episodes were made elsewhere and you only have the exported files, most things can still be measured, but a font you don't have can only be approached, not reproduced. A voice with no clean sample is hard to clone, and heavily compressed references soften colour measurements.
Common questions
Do I need the old project files? No. Vizard Agent measures from the exported episodes.
What if two episodes look different? Say which is right. Vizard Agent checks both first.
Why does the voice sound artificial? Usually the sample. Vizard Agent re-clones from a cleaner one.