The Artifact Atlas
StubbornIntrinsic to how these systems generate. Managed, not eliminated.

Identity drift

A subject's face, build or defining features change gradually over a clip, so the first second and the last second are recognisably different people.

Also called

  • Face morphing
  • Character inconsistency
  • Subject wander
Evidence

PLACEHOLDER EVIDENCE. Replace with a generation that shows a subject's features changing across the clip.

Causes
  1. 01Conditioning influence decays with distance from the anchor, so long segments end up effectively unanchored at the back end.
  2. 02Chained generation where each segment inherits from the previous one — small errors compound instead of cancelling.
  3. 03Identity carried in a text prompt rather than a reference, which underdetermines the face.
  4. 04Style conditioning strong enough to override identity conditioning; the look wins and the person loses.
  5. 05Reference imagery that is itself inconsistent, giving the model several faces to average.
What actually helps
  1. 01Shorten segments and re-anchor rather than asking one anchor to hold longer than it can.
  2. 02Anchor every segment to the same origin reference, never to its predecessor.
  3. 03Carry identity in reference imagery, not description.
  4. 04Check the balance between style and identity conditioning explicitly — they compete.
  5. 05Test at full intended length. Drift is invisible in a two-second test and obvious in a fifteen-second one.

Tends to originate at

  • Conditioning decay over sequence length
  • Chained generation
  • Weak identity anchoring

How to confirm it

Compare the first and last frames side by side, not the sequence. Drift is gradual by definition, which is exactly why watching it play makes it hard to see — each frame looks like a reasonable successor to the one before.

If the endpoints disagree and every adjacent pair looks fine, that is drift.

Why it is stubborn

Most artifacts on this list have a lever that reliably moves them. Drift is harder, because the mechanism causing it — conditioning influence falling off with distance — is intrinsic to how these systems generate, not a bug in a particular stage.

What is actually available is management rather than elimination: shorter segments, more frequent re-anchoring, and accepting a seam cost in exchange for identity stability. That trade is a real art-direction decision, not a technical detail, and it is worth making deliberately rather than discovering it by accident at the end of a project.