AI video face drift: diagnose and fix identity

When AI video face drift shows up, do not rewrite the whole prompt first. Classify within-clip morph, cross-clip mismatch, profile failure, or post-occlusion change—then fix references, shorten the clip, or lower motion. That path usually beats swapping model names.

Name the symptom first

Symptom	What you see	Change first
Within-clip	Early frames look right; later frames morph	Shorten duration; reduce camera/head turn; specify the end pose
Cross-clip	Each clip is fine alone; the edit string fails	Reuse one fixed reference set; never regenerate a “similar” face per shot
Profile / angle	Front holds; side or low angle collapses	Add a 3/4 reference; shrink the turn
After occlusion	Face changes after hands, hair, or props cover it	Run a no-occlusion baseline; then add one brief, shallow occlusion

For prevention workflows, see reference-frame consistency. This article is only about diagnosing drift that already happened.

Six-step triage

1. Regenerate once with identical settings
   Rule out chance. Two matching failures mean a systematic issue.

2. Audit the reference set
   Line the images up: hair part, glasses, makeup, apparent age. Drop outliers. Selection rules still follow how to pick reference frames.

3. Strip appearance words from the prompt
   With a first frame or identity refs, delete hair color, eye shape, and beauty-portrait style tags. Keep the five-part prompt structure, but hand looks to the image.

4. Simplify the shot
   Locked or tiny push; face toward camera; medium close-up so the face has enough pixels. Orbits, whip pans, and wide full-body shots are stress tests—not first-aid.

5. Shorten the clip
   If 10s drifts and 5s holds, cut two shorts and edit. Align motion with seconds in camera move and duration.

6. Confirm the input mode
   Need a specific face but still on pure text-to-video? Switch to first-frame or multimodal reference. Mode choice: text-to-video vs image-to-video.

Fixes by input mode

Mode	Common drift cause	First cut
Text only (t2v)	No visual identity anchor	Lock a still, then use first-frame / image-to-video
First frame (i2v)	Prompt rewrites appearance; motion too strong	Delete look words; lower action; rerun short
Start + end frames	Ends disagree on framing or makeup	Closer pair of stills; prompt only the transition
Multimodal reference (r2v)	Conflicting refs or multiple subjects	1–3 consistent refs of one person; name the primary

Rule of thumb: references own identity; prompts own motion. When both fight, the face fails first.

A copy-ready “stop the bleed” prompt skeleton

With a first frame, write motion—not looks:

Medium close-up, shoulders facing camera. She blinks once slowly, eyes drift toward the window, head turn under about ten degrees. Locked camera, soft side window light, calm expression. Ends still facing camera.

Avoid: perfect features, celebrity face, new hairstyle, 360° orbit while laughing—those invite drift on purpose.

Wrap-up

Face drift is a diagnostic problem. Name one of the four symptoms → run the six steps changing one class of variables → switch to first-frame or multimodal reference when text alone cannot hold identity. Polish style only after the face holds; before publish, skim the commercial checklist. To verify quickly, open the create page, pick the right input mode, and run a short test clip.