Why faces drift
Ask an image model for the same character twice and you get two cousins. Over a whole video that reads as a different person in every shot, and it is the quickest way to make a viewer stop believing the story.
One portrait, sent with every frame
When a video has a recurring lead, that lead is drawn once, as a reference portrait nobody ever sees. The portrait goes along with every frame the character appears in, and the prompt forbids the face, hair, age, build or outfit from changing between shots.
Places get the same treatment. Each location is drawn once, empty, and every scene set there is drawn with that plate behind it, so a lobby stays the same lobby.
A check before the clip is made
Before a frame is animated it is compared with the portraits and the place it should show. A changed face, a missing character or the wrong room gets it drawn once more with the problem named.
The video model sees only that first frame, so each motion prompt also asks it to keep the faces, clothes and place as they are, and asks one person for one action at a time.