How to Keep the Same Character Across AI Video Shots

One great AI video clip is easy. Four clips of the same character is where it gets hard. The hair changes length between shots, the coat changes colour, the face becomes a cousin of the face you started with. Viewers notice immediately, and the sequence stops feeling like a story.

Consistency isn't luck. It comes from a small system: fixed reference images, identical wording, and shots that don't ask the model to reinvent the character.

Why characters drift

Each generation is independent. The model doesn't remember your last clip. Every time, it rebuilds the character from whatever you gave it: the start image and the words. If the words change even slightly (“red coat”, then “crimson jacket”), or if the start image was generated separately with slightly different features, you get a slightly different person.

Drift also happens inside a single clip, when the character turns away, gets small in the frame, or moves fast. The model loses track of details and fills them back in with guesses.

Step 1: build a character card

Write one short description of your character and never change it. Paste it identically into every image prompt and every video prompt.

A woman in her thirties with short black bob hair, a long red wool coat, black boots, silver hoop earrings.

Five or six visual anchors: age, hair, one signature clothing item, colours, one accessory. Specific, visible, easy to repeat. Avoid anything that changes with angle or light, like “piercing eyes”.

Step 2: make every start frame from the same source

The strongest consistency comes from images, not words. Generate one clear, well-lit reference image of your character. Then make each shot's start frame from it, using your image tool's reference, edit or character features, rather than generating each one from scratch.

Every start frame should share the character card, the same lighting style and the same film look. Once you have the frames, the video prompt only needs to describe motion. Our image to video guide explains why re-describing the picture causes drift of its own.

Full-length studio photo of a woman with a short black bob in a long red wool coat, black boots and silver hoop earrings — A reference image made from the character card: full length, neutral backdrop, soft even light, so every anchor in the card is clearly visible. Each start frame is then made from this one picture. Made with GPT Image from a prompt built with Frame Juice Image settings: Photograph, Full Body, Eye Level, 50mm, f/4, Studio Soft, True Color.

The prompt behind this image:

sharp subject: a woman in her thirties with short black bob hair, a long red wool coat, black boots, silver hoop earrings, standing and facing the camera, blurred background: a plain light grey studio backdrop, real photograph captured with a camera, full body shot, head to toe framing, eye level camera angle, neutral natural perspective, 50mm standard prime lens, mirrors human eye perspective, neutral compression, timeless rendering, f/4 aperture, slight background defocus, context visible, large softbox studio lighting, soft wraparound light, controlled shadows, commercial product quality, true to life color rendering, neutral color grade, accurate skin tones and whites, natural look, professional photography, highly detailed, photorealistic.

Step 3: a location card too

The world drifts as much as the character. Write a location card the same way.

A narrow rainy night market with red paper lanterns, wet cobblestones and steam from food stalls.

Same wording, every shot.

Cards that never change

Frame Juice Motion Pro's shot list keeps a locked character card and location card that are pasted identically into every shot, with an identity lock setting for each prompt.

See Frame Juice Motion

Step 4: shots that protect identity

Keep the face visible. Long stretches with the character facing away invite the model to rebuild the face when they turn back.
Avoid tiny characters. A face that's twenty pixels wide in a wide shot has no details to hold on to.
Change angles in steps. Cutting from a front close-up to a top-down wide is a lot to reconcile. Go from medium front, to three-quarter close-up, to over the shoulder.
Slow actions. Fast spins, running and jumping blur details and invite drift.
Stable rigs. Tripod, dolly and gimbal moves keep the character readable. Chaotic handheld makes it harder.

Camera on a dolly track pushing in toward a woman seated at a kitchen table, with open space around her — Slow Push-In Drama, a camera move in Frame Juice Motion. A slow move on a stable rig, with the face in view, gives the model little reason to rebuild the character.

Step 5: ask for identity to stay locked

A short line in each video prompt helps:

The woman's face, hair and clothing stay exactly the same throughout.

On platforms with a negative prompt field, add changing identity, face morphing there as well.

A four-shot example

Shot 1: wide, the woman enters the market, lanterns overhead. Slow tracking shot from behind.

Shot 2: medium, she stops at a noodle stall. Static camera, steam rising.

Shot 3: close-up, she looks up at something off-screen. Slow push-in.

Shot 4: over the shoulder, we see what she sees: a figure in the crowd. Static, rack focus to the figure.

Each shot gets the same character card, the same location card, a start frame from the same reference, and one move. Written out, shot 3 looks like this:

Slow push-in. A woman in her thirties with short black bob hair, a long red wool coat, black boots, silver hoop earrings, in a narrow rainy night market with red paper lanterns, wet cobblestones and steam from food stalls. She slowly looks up at something off-screen. The woman's face, hair and clothing stay exactly the same throughout.

Close-up of a woman with a black bob, hoop earrings and a red coat looking up in a rainy night market lit by red lanterns — A frame for shot 3, made from the character card and location card words alone. The coat, bob and earrings match, but the face is a different woman from the reference above. That is drift, and it's why step 2 makes every start frame from one reference image. Made with GPT Image from a prompt built with Frame Juice Image and Motion settings: Cinema, Cinematic Drama, Dolly on Track, Push-In, 85mm, Subtle Life, Smoke and Steam, Close-Up, f/1.8, Cinestill 800T.

The prompt behind this image:

sharp subject: still frame from a slow push-in, a woman in her thirties with short black bob hair, a long red wool coat, silver hoop earrings, slowly looking up at something off-screen, blurred background: a narrow rainy night market with red paper lanterns, wet cobblestones and steam from food stalls, cinematic motion picture still, 24fps filmic look, cinematic drama pacing, restrained motivated camera work, camera on a dolly track, perfectly smooth linear travel, the camera pushes in toward the eyes of the subject, 85mm lens, flattering compression with a soft background, subtle living motion only: breathing, blinking and tiny natural movements, smoke and steam rising and curling, close-up shot, face filling the frame, f/1.8 wide aperture, beautiful bokeh, smooth background separation, Cinestill 800T tungsten cinema film stock, iconic red halation glow, cinematic night atmosphere, warm tungsten color, professional photography, highly detailed, photorealistic.

Camera crew on a dolly track following a man in a trench coat and fedora down a foggy, rain-wet street at night — Noir Tracking, a camera move in Frame Juice Motion. A smooth tracking shot keeps the character a readable size in the frame, which suits an opening shot like shot 1.

When it still drifts

Regenerate only the drifting shot, not the whole sequence. If one angle keeps failing, change the angle. And in the edit, cuts hide small differences: a quick cut away and back is far kinder to consistency than one long continuous shot.

If you need a shot to end on a specific pose or place, give it an end frame from the same reference as well. Our start and end frame guide shows how.

How to Keep the Same Character Across AI Video Shots

Why characters drift

Step 1: build a character card

Step 2: make every start frame from the same source

Step 3: a location card too

Step 4: shots that protect identity

Step 5: ask for identity to stay locked

A four-shot example

When it still drifts

Make your AI clips move with intent.

More articles

Veo vs Kling vs Runway vs Seedance vs Wan: Prompt Formats Compared

Image to Video Prompts: What to Write When the Image Sets the Look

Camera moves for AI video: a working vocabulary