The most common complaint about AI video is not quality, it is identity. Your hero has brown eyes in shot one, a different jawline in shot two and a new jacket in shot three. Each clip looks fine alone; together they look like different people.
Consistency is mostly a workflow problem rather than a model problem. If you decide the character once, feed that decision into every shot and check it cheaply before you render motion, drift drops sharply. This guide walks through that workflow. Larund, the studio we build, has it built in, but every step also works with other tools.
Why characters drift between AI video shots
A video model generates each clip from scratch, guided by your prompt and, if you supply it, a starting image. It does not remember the previous clip. Anything you leave unspecified, such as the exact hairline, the colour of a collar or the shape of a nose, is filled in with a plausible guess, and the guess changes every time.
Drift gets worse in three situations: when the character turns away from the camera, when the shot is long and the face spends many frames at an angle the model handles poorly, and when you mix several models, each of which draws faces slightly differently.
So the goal is to remove guesses. Give the model one fixed visual description, and ideally one fixed image, every time it draws your character.
Step 1: design the character once
Before any video, make the character as a still image. Stills are cheap, fast to iterate and easy to compare side by side. Choose an image model that handles faces well, generate several options and pick one.
Write a short fixed description that you will paste into every later prompt. Include the traits that carry identity: age range, face shape, hair colour and style, skin tone, distinguishing marks, and the default outfit. Avoid vague words like “handsome” and prefer concrete ones like “short curly black hair, round glasses, olive green bomber jacket”.
- Generate a neutral front-facing portrait on a plain background, good lighting, relaxed expression.
- Add a three-quarter view and, if the character will walk away or be filmed from the side, a side or back view in the same outfit.
- Write the fixed description (about 30 to 50 words) and save it as a text note.
- Pick one image as the master reference and keep the others as supporting angles.
Step 2: lock the world, not only the face
Identity is not just the face. Lighting, colour palette and location change how a face reads, and a character who looks different in daylight and neon is often judged as inconsistent. Decide the look of each location once and reuse it.
Larund formalises this as two sheets before any video: a location sheet for the place and a cast sheet for each character. Every shot in the scene is then built from those sheets rather than from a fresh description. If you work elsewhere, create the same thing by hand: one reference image of the room or street, one of each person, and a fixed style line such as “soft overcast daylight, muted teal and orange grade”.
Step 3: approve keyframes before you pay for motion
A keyframe is a still of the first or last moment of a shot. If the still already shows the wrong face, the video will too, so keyframes are the cheapest place to catch drift.
In Larund each shot gets a start keyframe and an end keyframe generated from your cast and location sheets, and you approve them before motion is rendered. The credit price is shown before every generation, so a bad keyframe costs you a still, not a clip.
- Create the start keyframe using the master reference and the fixed description.
- Check the face, outfit, props and background against your reference. Regenerate if anything is off.
- Create the end keyframe the same way, showing where the character ends up.
- Only when both frames look right, render the motion between them.
Step 4: write prompts that protect identity
Keep the motion prompt about motion. The look is already in the keyframes, so use the prompt for action, camera and mood, and repeat only the short identity description. Contradicting the reference in the prompt, for example asking for a different hairstyle, invites the model to redraw the face.
Prefer small, readable actions in each shot: a head turn, a step forward, picking up a cup. Large changes of pose, fast spins and heavy occlusion are where faces break. Lens and angle tokens help too; in Larund, tokens such as &35mm or &low-angle pin the shot so the framing stays predictable across scenes.
- Keep shots short; consistency is easier over a few seconds than over a long take.
- Prefer front and three-quarter angles for close-ups.
- Do not stack several people with similar looks in one prompt; give each a distinct outfit colour.
Step 5: choose models deliberately
Different video models differ in how closely they follow a start image and how they handle faces. Some are strong at realistic human motion, some at stylised animation. Pick one model per scene and stay with it inside that scene, because switching mid-scene is a common source of visible changes.
Because Larund keeps one credit balance across 25+ video models, you can test the same keyframes on two models and keep the one that holds the face better. Models are chosen by their real names, with samples, so the choice is visible rather than hidden behind a default.
| Problem | Likely cause | Fix |
|---|---|---|
| Face changes after a head turn | Shot too long or angle too extreme | Split into two shots with a new keyframe |
| Outfit colour shifts | Outfit not in the fixed description | Add the exact colour and garment to every prompt |
| Character looks different between scenes | Different reference or model | Reuse the master reference and the same model per scene |
| Two characters blend | Similar looks in one frame | Make outfits and hair clearly different |
When to reach for a dedicated character feature
If your character is a recurring persona rather than a one-off film role, such as a brand ambassador who appears in weekly posts, treat the character as an asset. Larund’s AI Influencer feature places one consistent character in new scenes, which suits that case better than rebuilding the character each time.
For narrative work with several characters and locations, a scene-by-scene tool that plans a shot list and chains scenes is a better fit. Larund’s Cinema Studio does that, and the same approach applies in any tool: define cast and locations first, then shots.
Is Larund the right pick?
A good fit when
- You need the same character in many shots, such as a short film, a series of ads or a recurring brand mascot.
- You want to approve how a shot looks before paying for the motion.
- You want to switch video models between scenes without losing the character.
Probably not the best fit when
- You only need one single clip with one character; a plain image-to-video prompt is enough.
- You already have a trained, custom character model in another tool and want to stay inside it.
- You need frame-by-frame editing and compositing on a desktop editor timeline.
Frequently asked questions
How do I keep the same character in every AI video shot?
Create one master reference image, reuse it as the starting image for every shot, repeat a short fixed description of the character in each prompt and approve a still keyframe before rendering motion.
Can I use the same character across different AI video models?
Yes, if you start every shot from the same reference image. Results vary by model, so test your keyframes on a couple of models and stay with one per scene. In Larund one credit balance covers 25+ video models.
Why does my AI character’s face change when they turn their head?
The model must invent the unseen side of the face and does so differently each time. Use shorter shots, supply a three-quarter or side reference and split large turns into two shots.
Is there a free way to try character-consistent AI video?
Larund has a Free plan with 5 credits a month plus 20 one-time welcome credits, roughly 4 short videos or 28 images. Free videos carry a small larund.com watermark.
Do I need a trained custom model to keep a character consistent?
Not necessarily. For short projects, a good reference image, fixed wording and approved keyframes usually go a long way. Custom training can help for large, long-running projects but takes more setup.
How many reference images should I make for one character?
Usually three: front, three-quarter and side or back, all in the same outfit and lighting. One strong front-facing image is enough for simple close-up shots.
What is a cast sheet in AI video?
A cast sheet is a reference image set for a character showing consistent face and outfit from several angles. In Larund, each scene is built from a cast sheet and a location sheet so every shot reuses the same look.
Every top model. One studio.
Start free with Larund Studio: 25+ video and 30+ image models, keyframe films and motion graphics, with the price shown before every generation.
Try Larund free