AI Video Faces Look Wrong: Why It Happens and How to Fix It in Your Prompt
The specific reasons AI-generated faces distort, drift, and degrade — and the exact prompt techniques that prevent each one.
If you've generated enough AI video, you've seen it: a face that looks perfect in the source image slowly melts into something uncanny over the course of a few seconds. Eyes that drift apart. A jaw that subtly reshapes. Skin that develops an oily, plastic quality that reads as fake immediately.
Face distortion is the most common complaint from AI video creators — and almost all of it is avoidable with the right prompt technique.
Here's what's actually happening and how to fix it.
Why faces distort in AI video
AI video models don't "understand" faces the way humans do. They treat faces as a region of pixels that should remain roughly consistent across frames while also changing in ways that match the motion described in the prompt.
The problem is that faces are the thing humans are most sensitive to. We're biologically wired to detect subtle wrongness in faces — it's why the uncanny valley effect is so powerful. A model that does a perfectly acceptable job on a tree or a building will still look wrong if it introduces a 2% deviation in the distance between someone's eyes.
Face distortion happens for four main reasons:
1. The model has too much to animate at once
2. The clip is too long for the model to maintain consistency
3. The camera is too close for too long
4. The prompt is asking for expressions the source image doesn't support
Each has a specific fix.
Fix 1: Give the face less to do
The more motion you ask for near a face, the more the model has to recompute facial geometry across frames — and the more opportunities it has to get it wrong.
Bad prompt:
Woman laughing and turning her head while wind blows her hair, camera slowly pushes in toward her face.
That's three simultaneous motion systems on or near the face: expression change (laughing), head rotation, and hair physics. Combined with a push-in camera that makes the face larger in frame over time, this is a recipe for distortion.
Fixed prompt:
Woman with a slight smile, head still. Wind moves her hair gently. Camera holds static.
One face state (slight smile, already present in source image). One motion system (hair). Static camera. The face barely has to move — and therefore barely has a chance to distort.
Fix 2: Keep clips short
Every AI video model has a consistency horizon — a clip length beyond which it can no longer reliably maintain facial geometry. For most current models, this is around 4-6 seconds for close-up face shots.
Runway Gen-4 is particularly prone to face drift after the 4-second mark. Kling handles slightly longer clips but starts to lose fine detail (eyelashes, skin texture) around 6-7 seconds.
The fix: For any shot where a face is prominent, target 3-4 seconds maximum. If you need a longer clip, generate two shorter clips and cut between them. The edit will look more intentional than a single long clip with a drifting face.
Fix 3: Watch your framing
The closer the camera is to a face, the more visible any distortion becomes. A 2% deviation in eye spacing is invisible in a wide shot and obvious in a close-up.
Framings from safest to riskiest:
- Wide shot (person small in frame) → very safe
- Medium shot (person waist-up) → safe
- Medium close-up (chest and above) → acceptable
- Close-up (head and shoulders) → risky, keep short
- Extreme close-up (face fills frame) → avoid for more than 2 seconds
If your prompt requires a close-up, keep it static. A push-in that ends in extreme close-up is one of the fastest ways to reveal face distortion.
Bad prompt:
Slow push in from medium shot to extreme close-up of woman's face.
Fixed prompt:
Medium close-up of woman, camera holds static. Slight smile, no other movement.
Fix 4: Don't ask for expressions that aren't in the source image
This is the one most people miss. If your source image shows a person with a neutral expression and your prompt asks them to smile or laugh, the model has to generate a new expression from scratch — interpolating between neutral and smiling across frames. This process almost always produces distortion.
The model is much better at *maintaining* an expression than *creating* one.
Bad prompt (with neutral source image):
Woman breaks into a wide smile as she looks up from her book.
Fixed prompt (with neutral source image):
Woman reading, slight downward gaze, expression calm and focused. Page turns slowly.
Or better: if you need a smiling face, *generate the source image with the smile already present*. Then your video prompt just needs to maintain it, which models do reliably.
The rule: Your source image expression is your ceiling. Don't ask the video generation to go beyond it.
Which models handle faces best
Not all models are equal on this dimension:
Kling — Strong face consistency for still or minimally moving faces. Struggles with expression changes and extreme close-ups over 5 seconds.
Runway Gen-4 — Good face detail retention but drifts on longer clips. Use motion brush to isolate face region if you need precise control.
Seedance — Excellent for faces in motion-heavy scenes (wind, action) because its physics engine handles surrounding elements well, reducing the motion burden on the face itself. Less ideal for sustained facial close-ups.
Hailuo (MiniMax) — Arguably the strongest face consistency of any current model for portrait-style shots. Worth testing specifically for close-up face content.
The combined approach
For any shot where a face is the primary subject:
1. Generate your source image with the target expression already present
2. Keep the camera at medium close-up or wider
3. Give the face minimal motion — breath, slight head tilt at most
4. Target 3-4 seconds
5. Use Hailuo or Kling for portrait shots; Runway for scenes with more environmental context
Before you generate, run your prompt through Dry Run. The image dependency score will tell you how much your result depends on getting the source image right — and the suggestions will flag specific prompt elements likely to cause face distortion before you spend credits finding out.