Your AI character looks right in the opening frame, then their eyes, jawline, hair or accessories shift as the video plays. This is identity drift. A face turn, camera movement or momentary occlusion can expose details the generator cannot maintain consistently.
Quick answer: Use a clear character reference when your model supports one, start with a short fixed-camera shot and restrained head movement, and compare the opening, middle and ending frames. If a face still morphs, simplify the motion or split the scene into separate shots. A prompt alone cannot guarantee identity consistency.
Our real five-second AI video: the source and settings
On November 10, 2026, AI Revolution generated the original clip below with Seedance 2.0 Mini accessed through Runway. It is an actual generated video, not stock footage or a fabricated before-and-after comparison. The clip was made as a head-turn consistency example, not a controlled benchmark.
Our five-second Seedance 2.0 Mini head-turn generation example, November 10, 2026. This is a generation example, not a controlled identity-drift benchmark. Watch on YouTube.
Exact generation prompt:
A single fictional adult woman with short dark curly hair, a distinct small silver star-shaped earring in her left ear, wearing a mustard yellow raincoat, stands in a softly lit studio with plain gray background. One continuous locked-off medium close-up. She slowly turns her head from front view to a three-quarter view and back, then smiles gently. Natural realistic motion, stable identity, stable facial proportions and unchanged clothes, no cuts, no other people, no text.
Recorded settings: Seedance 2.0 Mini; 16:9; 720p; five seconds; audio disabled. Recorded attempts: one clip is documented here. No separate starting reference image is documented for this generation, so this example should not be treated as proof of reference-image performance. The credit cost was not recorded and is not estimated.
What we observed: In the five-second clip, the woman begins with a neutral expression, turns her head to one side and then the other, returns to face forward and smiles. In the owner’s normal-speed viewing, she remains recognizable and no obvious major face morphing was noticed. This is a qualitative viewing observation, not a verified frame-by-frame comparison of facial proportions. We have not measured subtle changes in the jaw, eyes or earring.
Featured image note: The male face-morphing cover is an AI-generated editorial illustration, not a sequence of frames from the woman in our real video experiment.
Why faces change during AI video generation
A video generator has to produce a plausible face across changing angles and frames. When the head rotates, the model must infer facial surfaces that were not visible in the initial view. Hair, earrings, teeth and hands passing in front of the face add occlusion; fast camera motion and lighting changes make visual comparison harder. A stable-looking first frame is therefore not enough. Generative video models also have to maintain temporal consistency across changing views. They do not necessarily maintain an explicit, fixed 3D identity model, so facial details can vary between frames. The exact generation process differs by model; it is not accurate to assume every system simply predicts the next frame from the previous one.
There are also two different problems: within-clip drift, when a face changes during one shot, and between-clip inconsistency, when a new generation produces a different-looking character. The first calls for motion and frame inspection; the second often benefits from reusing the same character references. For a broader production workflow, read our guide to keeping one AI character across multiple videos.
How to reduce identity drift: a practical workflow
- Lock the character design. Choose a sharp, evenly lit portrait with a visible face and recognizable features. Prefer a clearly visible, reasonably front-facing face over a distant, heavily obscured or strongly shadowed portrait; this is a practical starting point, not a measured guarantee. For repeated scenes, save a front and three-quarter view if your tools support multiple references.
- Use the reference control actually supported by your model. Some models accept a starting image, some accept character/reference images, and some are text-only. Do not assume a reference was used just because the prompt says “same character.”
- Start with an easy shot. Use a fixed camera, medium close-up, subtle expression and a small head turn. For example, test a gentle three-quarter turn instead of a full 90-degree profile turn. Avoid simultaneous camera orbit, rapid rotation and hand-over-face motion.
- Describe the action, not just the desired outcome. State the camera position, the exact movement, wardrobe and one or two distinctive details. “Stable identity” can be included, but it is not a hard constraint.
- Inspect the whole five-second clip. Pause near the start, midpoint and end. Check eye spacing, jaw, hairline, earrings and jacket seams. Compare like-for-like angles where possible.
- Change one variable at a time. If the face shifts, first reduce rotation; then try a stronger reference or shorter shot. Record each prompt, model and setting so you know what changed.
Copyable prompt for a restrained head-turn shot
Use the same fictional adult woman shown in the supplied character reference. Medium close-up, locked camera, soft even studio lighting. She makes a small, slow head turn from front-facing to a gentle three-quarter angle, pauses, then looks forward again. Preserve her facial proportions, hairline, eye shape, distinctive earring and clothing. One continuous shot, subtle natural motion, no cuts, no additional people.
Use the opening sentence only if you actually attach a character reference to a compatible model. For a text-only generation, replace it with a full visual description. This is a suggested prompt, not the prompt used for our embedded clip.
Which fix should you try first?
| What you see | First change to test |
|---|---|
| Face changes during a head turn | Try a smaller three-quarter turn instead of a full profile; keep the camera still |
| Hair or earring changes after occlusion | Remove the occlusion or use a clearer reference |
| New clip looks like another person | Reuse the same character reference, not just the text prompt |
| Face changes after extending a clip | Inspect the transition frame; consider a cut to a fresh short shot |
| Face looks different only because of lighting | Compare frames at similar angles and lighting before declaring drift |
How to check identity without inventing a success score
Save frames near 0 seconds, 2.5 seconds and the end of a five-second clip. Put them side by side at the same display size. Check whether the person remains recognizable, whether accessories appear or disappear, and whether the jaw and eye positions change in a way that cannot be explained by perspective. Note the timestamps of any problem. A three-frame contact sheet is a quick screening tool, not a substitute for watching the complete motion.
For a fair comparison between two prompts or models, use the same reference image, target action, aspect ratio and duration where the platforms permit. Repeat trials before making any claim about reliability; record multiple trials before drawing conclusions about reliability.
When to use a new shot instead of another regeneration
If a dramatic turn or camera orbit repeatedly changes the face, consider a simple cut between a frontal and a side-angle shot. A separately generated shot can look more intentional than forcing a single unstable rotation. If you need more control over how a clip starts and ends, see our first-and-last-frame video guide; these controls are model-dependent and do not guarantee the identity in every intermediate frame.
Official guidance and limitations
Runway recommends clear character references and reusable character views for consistency across shots; see its character consistency guide and guide to character plates for longer videos. Reference features differ by model and product. Midjourney’s video documentation likewise distinguishes lower and higher motion, noting that high motion can introduce glitches. These are documented workflows, not measurements of our particular clip.
The takeaway
Treat face morphing as a shot-design and verification problem, not something a single magic prompt can solve. Start with a good reference where supported, reduce simultaneous motion, inspect the full clip and preserve your exact settings. Our embedded video is one documented generation example, not a model reliability benchmark. If complex motion repeatedly changes a face, try several shorter, simpler shots instead of forcing one long rotation.