How to stop an avatar making faces the real person never makes
Tell Vizard Agent the face is right but the expressions are not. It goes back to your source video, studies the range the person actually uses, maps where their face is unobstructed second by second, builds a base cut from that clean footage and lip-syncs it to the new script. The expressions are then theirs rather than invented.
What do you type into Vizard Agent?
Describe the person rather than the fault. Saying he never pulls those faces and pointing at the original gives Vizard Agent evidence it can study, while "make it more natural" is a direction with no target. Ask for their own footage lip-synced to the new script, because that's the method that fixes it.
Prompt
What does Vizard Agent actually do?
It stops generating a face and uses the one already on file. A video that seems to show someone talking for two minutes usually holds far less usable face than that, because a hand goes up or they look away. So the base cut is built only from footage that survives a visibility check. In the session behind this article it:
- Took a spokesperson video whose generated avatar performed far more than its subject ever did, and studied the real expression style across the original footage.
- Mapped where the face was unobstructed, checking visibility second by second, then assembled a base cut from the clean spans and reviewed it before syncing anything.
- Lip-synced the real footage to the new voice track, then measured where the head sat in each segment, since spans from different parts of a recording put it in a slightly different place.
- Reframed the base cut tighter on the face, lip-synced the reframed footage and zoomed into the mouth region to check the sync closely.
- Finished with captions and music over the new cut and had the finished version watched back end to end.
When does this not work well?
You need clean footage of the face. If the source is short, badly lit, mostly in profile, or has something across the face for most of its length, there may not be enough to build a base cut from. Then a generated avatar really is the only option. A long new script on short source footage also means spans have to repeat.
Common questions
How much source footage is needed? More than you expect, because Vizard Agent counts only clean face time.
Why not just prompt for calmer expressions? A description of somebody's manner is a guess. Their own footage, which Vizard Agent uses, is a record of it.
What if I only have photographs? Then a generated performance is the only option Vizard Agent has.