Keep one character consistent across several AI video clips

Every clip you generate is a fresh roll of the dice. The model does not remember the courier in the mustard raincoat from the clip before; it only knows what you give it this time. So when shot two shows a coat that is a little yellower and a face that is a little rounder, nothing is broken. You simply gave the model room to guess.

Consistency is therefore not a setting. It is a habit of giving the model the same information, in the same words and the same pictures, every time, and of knowing in advance which details it will still get wrong.

Start with one fixed reference still

Words describe a character loosely; a picture pins it down. Make one clean still of your character first: full body or waist up, plain background, even light, facing the camera or a little to the side. Fix that picture before you spend anything on motion. A text-to-image model such as the still-image models in the catalog is fine for it, and an edit model like Nano Banana 2.1 edit lets you change a coat colour or a background without redrawing the person.

Then start every clip from that still on an image-to-video model such as Kling V4.0 image-to-video. The first frame fixes the face, the clothes and the palette; your prompt only has to say what happens. A text-only clip invents the character again each time. The workflow in draft, lock, finish works well here: draft the motion cheaply from the still, then spend on the final render once.

Write the character line once, then paste it

Next to the picture, keep a single sentence that names the character. Write it once, in the order you like, and paste it unchanged at the start of every prompt. Do not rephrase it for variety. "Mustard raincoat" in one prompt and "yellow jacket" in the next are two different coats to a model.

• Fix the nouns and the colour words. Choose one name for each garment and one colour word, and never swap them. Specific beats poetic: "mustard" is better than "golden".
• Keep it short. Five or six traits is plenty: age range, hair, one main garment, one accessory, one colour. A long list gets partly ignored.
• Put the character first, the action second. The line is the fixed part; only the action and the camera change from clip to clip.
• Leave out what you do not need. Every extra detail is one more thing that can change between shots.

A character sheet beats a memory

Before the second clip, write the sheet down: the reference still, the character line, the palette in three colour words, the light words and the shape of the shot (tall or wide). Keep it in the same note as your shot list, so you can paste from it instead of from the last prompt you remember. A shot list gives the order of the clips; the sheet gives each of them the same cast.

If you have a few extra stills of the same character from other angles, a side view and a back view help most, keep them with the sheet. Models that accept several reference images can use them; on a single-image model they are your own check, so you can compare a result with them.

Open on the frame the last clip ended on

When two clips follow each other in the cut, the second one should begin where the first one stopped. Open the finished clip in a video editor, step to its final frame, export that frame as a still and use it as the first frame of the next clip. The pose, the position in the room and the light are already right, and the model only has to continue.

• Pick a calm last frame. A frame caught mid-blink or mid-step gives the next clip a strange start. Step back a few frames to a steady one.
• Look at the exported frame before you use it. If the face has softened or the hand has gone odd at the end, the next clip inherits it. Choose an earlier frame, or go back to the reference still.
• Do not chain forever. Each generation from a generated frame loses a little. After two or three links, restart from the original reference still and cut the clips together.
• Pin both ends when you can. Vidu Q3 start-end-to-video and the last-frame option on Kling V4.0 let you say where a shot begins and where it ends, which keeps a short shot on one character from start to finish.

Keep the light the same from clip to clip

A character can be perfect and still look like a different person if the light moves from warm window light to cold overhead. Use the same two or three light words in every prompt of a scene, and change them only when the story changes the place. "Soft grey overcast light" stays on the street, on the stairs and in the hall. A cut from outside to inside is the moment to change it, and the cut hides the jump. If the light has to change inside one clip, the words in light changes without flicker keep it smooth.

Keep the frame shape the same too. A character drawn for tall and then asked for wide is composed anew, and the face and clothes tend to shift with it.

What still drifts, and how to cut around it

Even with a still, a pasted line and matching frames, some things move. Know them, and plan the shots so they matter less.

• Hands. Finger counts and grips change from clip to clip. Keep hands small in the frame, in pockets, behind an object or out of shot, and cut away before they do anything fine.
• Hair. A hairstyle gains or loses length between takes, and a ponytail turns into loose hair when the head turns. Pick hair that is easy to repeat: short, tied back or under a hood.
• Small logos and patterns. Badges, stripes, printed words and buckles are redrawn each time, and text on clothes is never stable. Choose plain clothing, and add any mark in an editor afterwards.
• Jewellery and glasses. Earrings change sides and glasses change shape. Skip them unless the story needs them.
• Far-away faces. A face that is small in the frame is rebuilt from few pixels, so it varies more. Keep important faces in medium and close shots.

The editor's tools are the answer: cut on an action so the viewer's eye is on the movement, show the back of the head or a shoulder in the shots where a face would be risky, insert a close-up of an object between two takes, and keep clips short. A clip can run up to 30 seconds at the very most, but the shorter the clip, the less time there is for the character to wander.

What the studio cannot do

It cannot guarantee that two clips show exactly the same person. There is no saved character, no face lock and no shared memory between runs, and a seed on a model that offers one repeats a result for the same prompt and picture, not for a new scene. The reference still, the pasted line and your cuts do the work. When a take is close but not exact, keep the clip and cut away from the difference rather than paying for a perfect one.