You wrote something like a woman in her thirties, mid-sentence, looking straight at the camera, hit generate, and got back a clip that made you close the tab. The mouth moves on its own schedule. The eyes don’t quite land anywhere. The face is almost a face, which is worse than not being one.
If that’s the loop you’re stuck in, the fix isn’t a better prompt for the face. It’s to stop pointing the camera at it.
Frontal close-ups are the single worst thing you can ask a current video model for, and it has nothing to do with how good the model is. You’ve aimed the lens at the one subject your viewer has spent their entire life becoming an expert in.
The face is where every error lands at once
A mid-shot of someone walking through a market can be wrong in a dozen ways and still read fine. Fabric folds, background drift, the way a hand brushes a doorway — none of it has to be correct, just plausible. Faces don’t get that grace.
On a frontal close-up, three separate things have to agree simultaneously: the shape of the mouth against the sound, the direction and liveliness of the gaze, and the micro-movement of the whole face — blinks, brow, the tiny slackening around the jaw that makes someone look like they’re thinking rather than posing. Any one of those failing drops you into the uncanny valley, and the other two can’t save you. Drift that would be invisible in a wide shot becomes the only thing the viewer sees when the face fills the frame.
There’s a second problem stacked on top. When you ask for a talking face without an audio reference, the model is doing two jobs at once: deciding what a mouth would do, and deciding what the rest of the face does while it does it. The result is a face performing an approximation of speech — a mouth shaped by an average of mouths rather than driven by a sound. That’s exactly the texture your eye flags.
And a third: the close-up is the only shot where the viewer stares at one thing for several seconds. Every other shot lets the eye roam and forgive. This one doesn’t.
You’re not generating a face, you’re generating trust
Here’s the reframe that changed how I write prompts. The shot you actually want is rarely “a person’s face.” It’s a person telling me something, or a person feeling something, or a person about to make a decision. Presence. The face is only one of maybe eight ways to deliver that, and right now it’s the most expensive and least reliable.
So build a shot kit and start reaching into it by default. The prompt language for each is short — you’re not describing a person, you’re describing a camera position and what’s allowed in frame.
The back of the head, over the shoulder. Over-the-shoulder shot, subject facing away toward a rain-streaked window, slow handheld drift, warm interior light against grey outside. You keep the person, you keep the scale, you keep the emotion — “someone standing at a window in the rain” reads as a mood instantly. And the model has no face to ruin.
The profile. Turn them ninety degrees and most of the risk evaporates. A three-quarter or full side view gives you expression through posture and jawline without demanding pixel-accurate lip work. Side profile, subject speaking, soft window light, static camera, shallow depth of field — you can hold this for several seconds.
Hands. Hands are where character content actually lives: the mug being turned, the pen tapped twice, the phone set face-down. Close-up on hands folding a letter, no face in frame, cool daylight, slight camera push-in. The viewer will assemble the person from two shots of hands and one of a rumpled bed and never once ask to see their face.
The empty frame. An insert with nobody in it is not a pause — it’s a line of dialogue. A chair pulled out. A kettle steaming. A door left open. Cut these between two face-free people shots and you have a scene with a plot and no rendering risk at all.
**Off-screen voice. ** Write the dialogue as audio, then shoot the words against a face-free image. This is the highest-leverage move on the list, because once a real voice carries the performance, the visuals only have to be atmosphere. It’s also how documentary and most radio-with-pictures work, for the same reason.
Silhouette, backlight, reflection, shadow. If you need a human shape and a hint of face, take the face away with light. Silhouette against a blown-out doorway, subject turns slightly, no visible facial detail. A reflection in a shop window or a dark screen does the same job and looks intentional rather than evasive.
Occlusion. Steam, a doorway, a foreground plant, a blurred shoulder. Something between the lens and the face buys you forgiveness. Partial information is far easier to believe than complete information that’s wrong.
When the brief won’t let you hide the face
Sponsored content, a talking-head explainer, a brand where your face is the product — you can’t always route around it. Then you negotiate, and the negotiation is always about frame size and time.
Pull back. A medium-wide or wide shot where the face occupies a small part of the frame is dramatically safer than a close-up, because the viewer’s attention is shared with the body and the room. Say it in the prompt: medium wide, subject small in frame, full body visible, room in focus.
Give the face motion. Walking, turning, tilting, passing behind something — a face that’s moving is a face whose errors read as movement rather than distortion. Locked-off static portraits are the worst case; keep them brief.
Never hold a line in a single take. Cut on the turn, the breath, the gesture. Two seconds of talking face, then hands, then the reaction, then back. The audience never gets long enough to inspect, and the edit reads as energy rather than as a workaround.
Light half the face. A hard shadow across one side, a strong backlight, a subject half out of a doorway — contour and contrast do the work that pixels can’t, and it looks like a choice a cinematographer made.
Redesign the shot. If the point is the product, the shot is hands on the product. If the point is a claim, the shot is the face reacting to someone else saying it. A reaction shot is easier to fake convincingly than a speech.
The edit is a sync tool
Most lip-sync complaints I see aren’t really rendering failures — they’re editing failures. A viewer notices mismatch most when a shot holds past the moment they start looking for it. Cut before that. The cutaway isn’t a cover-up; it’s rhythm, and it’s how the problem disappears without anyone noticing it ever existed.
Work audio-first. Record or generate the voice track, then build picture against it. Writing a visual prompt and hoping the model invents a matching performance is asking the wrong component to do the hard part.
Why does only the mouth look off?
Because that’s the part with a real-world reference point. Your viewer knows what this sentence looks like on a mouth, and the model is producing what some sentence looks like. Everything else on the face is judged by feel; the mouth is judged by comparison.
Can I fix lip sync in post?
Sometimes, with tools that re-time mouth regions to audio — but you’re spending the effort on the riskiest shot in the piece. Take the five minutes and change the camera angle instead. It’s cheaper and it looks better.
Do talking-head videos actually need a face?
No. The most believable character content I’ve made this year has a voice, hands, a room, and one silhouette. The face was never carrying the story; the performance was.
What if my audience expects to see me?
Then you’re in a relationship, not a rendering problem. Show your face in a wide, well-lit, moving shot, keep the close-ups for the moments you’re not speaking, and put the words over hands and room. Regular viewers forgive a lot when the voice and the framing are consistent.
The thing to do next
Open your prompt file and delete every “looking directly at the camera, close-up.” Replace the first one with an over-the-shoulder and the second with hands. Watch how much of the uncanny feeling was never about the model at all.
I keep a small library of these shot skeletons — backs, hands, inserts, silhouettes — written out as reusable prompt blocks rather than a pile of my own notes, and I’ve been cribbing the structure from the video and image prompt sets collected over at Awesome Prompts. Borrow the format, keep your own shot list, and stop spending your afternoon re-rolling a face that was never going to hold up in the first place.