Keep the same face across every shot

The single most common complaint about AI video is this: the character looks right in shot one, slightly off in shot two, and like their cousin by shot three. Faces drift. The good news is that keeping a character consistent is a solved problem now — it just isn't a prompt problem. It's a reference problem.

Why the face drifts

Left to a text prompt alone, the model reinvents your character on every generation. "A young woman with dark hair" describes a million people, and you get a different one each time. To hold a face steady, you have to hand the model something fixed to anchor to — not a better sentence, but a picture.

There are three references that work, from simplest to most powerful.

A reference image

The baseline: feed the model a single clear image of your character and tell it that this person, this outfit, these props appear in the output. Most image and video tools now accept a reference image directly. This alone fixes most casual drift — same face, same clothes, new scene.

A pose (movement) sheet

For animation and dance, one still isn't enough — you need to control how the character moves. The trick making the rounds: generate a pose sheet first.

• Ask an image model for a grid — say a 4×4, sixteen numbered panels — of the same character in a sequence of poses, on a plain background, like an animation reference board.
• Then feed that sheet to a video model and tell it to follow the panels in order: "animate the character through steps 1–16 in sequence."

Because every panel is the same character, the motion stays on-model from the first frame to the last. Creators have used exactly this — a sixteen-panel sheet from an image model, animated by a video model — to produce clean dance clips that hold identity the whole way through.

A depth map of a real clip

The most precise option: take a real video whose motion you want, extract its depth map (a grey, shape-only version of the footage), and use that as the motion reference. Then:

• The depth map drives the choreography, timing and framing — one to one.
• Your reference image supplies who is in the shot.

So you can take a fifteen-second clip you like, keep its exact movement, and swap in your own character dancing it. It's how people make endless "cover" videos without reusing anyone's actual footage — the motion is borrowed as shape, the person is yours.

Identity, not pixels

Here's the subtle part the pros figured out. When you hand a model a razor-sharp close-up of a face to copy, it often tries to copy the pixels — and in video that can come out low-resolution or stiff, because the face gets locked to a small patch of the reference.

The better mental model: give the video model enough to recognize the identity, then let it re-render the face fresh each frame. Some creators deliberately keep the reference face loose — recognizable but not pixel-perfect — so the model rebuilds a clean, full-resolution face that still reads as the same person. You're preserving who it is, not stamping down exact pixels.

A trick: video models hold a face better than image models

One more move worth knowing. Video models are built to keep a face stable while things move — that's their whole job. So some people run a video model to edit a still image: generate a tiny clip, keep the face locked, change the outfit or background, and pull the single best frame back out as a picture. It's a roundabout way to edit an image without the face turning into someone else — and it works because holding identity through motion is exactly what these models are good at.

The short version

Stop trying to word your way to a consistent character. Pick a reference: an image for casual consistency, a pose sheet for controlled motion, a depth map for exact choreography. Give the model enough to know who it is, and let it re-render. The face stops drifting the moment you give it something to hold onto.