Zombie video face consistency: find where identity changes

Name the failure you can actually see

Identity drift means a person’s recognizable features change over time. A role swap means the survivor and zombie assignments reverse. Occlusion means a face is hidden, so you cannot fairly judge it in that frame. Treating all three as “bad likeness” makes the next experiment vague.

• Drift: compare nose, eye spacing and jaw before and after the change.
• Role swap: check whether the zombie styling moved to the other person.
• Occlusion: locate the first frame where the face becomes visible again.
• Contact distortion: inspect the boundary between cheek, shoulder and hand.

Keep a short frame log

Save the source references and review the result at normal speed first. Then inspect the first visible failure and a clean frame before it. Record what changed on screen rather than guessing at the model’s internal cause.

A consistent naming scheme lets you compare revisions without relying on memory. Avoid changing the source photo, camera angle and action all at once. If all three change, you cannot tell which adjustment helped.

Choose the smallest relevant experiment

For a role swap, rewrite the A/B assignment and keep the same labels in each shot. For a hidden face, change the framing or remove the obstruction. For a distorted hug, simplify the movement. These are hypotheses to test in a tool that supports those controls.

If you observe drift before any motion, return to the references: unclear faces and strongly mismatched angles may make your direction ambiguous. Do not assume that adding “perfect face” repeatedly will solve a specific visible defect.

• Change one variable and label the revision.
• Compare the same moment in both versions.
• Keep the revision only if it improves the visible problem without introducing another.
• Stop retrying when the available controls cannot address the failure; consider a shorter or different shot.

Know when editing cannot repair identity

A crop can remove a distracting edge; it cannot recover a face that has changed identity. A cut before defective frames may preserve a usable moment, but only if the remaining action still makes sense. Blurring the whole face defeats a story based on recognition.

Our preview has no generated-video output to diagnose and no verified likeness score. Use this checklist on footage you legitimately obtain elsewhere; provider-specific controls and their effectiveness must be checked with that provider.