A consistent AI character comes from process, not from a cleverer prompt. Design the character as still images first, build a reference sheet, feed those references into every image and video generation, lock the wardrobe and the lighting, and plan the shots so the model is never asked for something it does badly. Then check every shot against the sheet before anyone else sees it.
Why do AI characters drift between shots?
- No memory. Every generation is a fresh sample. The model does not know what it made a minute ago.
- Words describe a type. “Woman, thirties, dark bob, green coat” fits millions of people, and the model picks a new one each time.
- New angles are guesses. If the model has only seen the face from the front, it invents the profile.
- Small faces lose detail. In a wide shot the face is a few dozen pixels, and the features get averaged into someone generic.
- Motion erodes identity. Within a clip, the face turns, gets covered, comes back, and returns slightly different.
- Light and style shift the person. Warm evening light or a change of grade can alter apparent age and skin tone.
- Errors compound. Using the last frame of one clip to start the next passes every small mistake down the line.
The image-first workflow for a consistent AI character
- Design one hero portrait. Iterate on a single still until the face is right. This is the cheapest stage at which to change your mind.
- Build the reference sheet from that portrait (the next section lists what goes on it).
- Make a still keyframe for every shot. All twenty, generated with the sheet as reference, in the right location, outfit and light.
- Approve the stills as a set. Lay them out together and fix the odd ones out.
- Animate each keyframe with image-to-video, adding character references where the model accepts them.
- Edit, grade and check.
The logic is simple. A video model that starts from a frame where the character is already correct only has to keep them correct for a few seconds. A video model that starts from text has to invent them again. It is the same order of work we describe in how AI reels are made and in our step-by-step reels workflow.
What goes on a character reference sheet?
- front, three-quarter and profile views of the head, from both sides if anything is asymmetrical (a parting, a mole, a scar)
- a full-body view for proportions, and a back view
- four to six expressions: neutral, smiling with teeth, mid-speech, surprised, concentrating
- hands, if they will be seen in close-up
- every outfit, with exact colours, and every accessory
- neutral, even light and a plain background, so that no mood gets baked into the references
- a short written spec: age, build, distinguishing marks and what never changes
- a version number, because there must only ever be one approved sheet
Which tools have character reference features in September 2026?
Most current models accept reference images. The details below come from each vendor’s own documentation as we checked it in September 2026. They change often, so check again before you plan around them.
| Tool | Feature | What the vendor documents |
|---|---|---|
| Google Nano Banana 2 and Nano Banana Pro (images) | Multiple reference images | Up to 14 reference images in one request, of which up to 4 (Nano Banana 2) or 5 (Pro) can be character images |
| Google Veo 3.1 (video) | Reference images, called “Ingredients to video” in Flow | Up to three reference images of a person, character or product, for 8-second clips |
| Kling 3.0 (video) | Element Library | An element is built from two to four images, a front-facing main image plus other angles, or from a 3 to 8 second video. A voice can be bound to a character element |
| ByteDance Seedance 2.0 and 2.5 (video) | Multimodal references | 2.0 accepts up to 9 images, 3 video clips and 3 audio clips. ByteDance says Seedance 2.5 accepts up to 30 images, 10 video clips and 10 audio clips |
| Runway (images) | Gen-4 Image References | Up to three active references per generation |
| Midjourney (images) | Omni Reference | One reference image set with --oref, weighted from 0 to 1000 with --ow. It runs on the V7 model: when V8.1 became the default in June 2026, Midjourney said V7 Omni Reference stays available while a V8 version is trained |
Two cautions. These features reduce drift and do not remove it. ByteDance’s own launch notes for Seedance 2.0 said that multi-subject consistency still had room to improve. And more reference slots do not mean better results: two or three well-chosen images that match the shot’s angle usually beat a dozen that pull in different directions.
When is a trained custom model worth it?
A LoRA is a small add-on model trained on a set of images of one character and used with an open-weight image or video model. Some platforms sell the same idea as a built-in “trained character” feature. Training takes more preparation than uploading references, so it has to earn its place.
It is worth it for a series that will run for months, for a brand mascot or a lasting AI presenter, for stylised animated characters that reference features read poorly, and for projects where several people generate the same character. It is rarely worth it for one 20-shot film.
An invented character has no photographs to train on, so the dataset has to be generated with reference features first and then curated hard. Every flawed image you keep teaches the flaw. If the character is a real person, written consent comes before any training, and the agreement should say when the model is deleted. We cover that side in our guide to AI spokesperson videos. If the real person is you, our piece on AI avatar videos for founders explains what to record and how voice cloning fits in.
How do you keep wardrobe and lighting continuous?
Wardrobe. Use one outfit per scene and describe it with the same words every time. Keep a “character block” of text and paste it, never retype it. Choose solid colours and a strong silhouette, and avoid fine patterns, small logos, lettering and delicate jewellery, because fine detail mutates between shots. A product in the character’s hands drifts for the same reason, and we explain the fix in our guide to keeping the product accurate in AI product videos. A distinctive coat does a surprising amount of identity work: viewers register shape and colour before they register a face.
Lighting. Define it per scene: direction, colour temperature, time of day. Generate all the keyframes for one scene in the same session with the same environment reference. Leave the final look to the grade, where one adjustment applies to every shot equally.
A continuity log. Keep a spreadsheet with a row per shot: outfit, hair, props, light, which references were used, which model and version. Do not change model version halfway through a project unless you are prepared to redo earlier shots.
How do you plan shots to hide the weaknesses?
- Put the shots that depend on the face in medium and close framing, where references work best.
- Make wide shots properly wide, so the character reads as a silhouette. The risky distance is the one in between, where the face is visible and too small to hold.
- Keep clips short and cut before drift sets in.
- Avoid full head turns, hands passing across the face, and characters who leave the frame and come back.
- Use cutaways that do not show the face: hands, objects, over-the-shoulder views, the back of the head.
- Keep to one character per shot where the story allows. Two-shots are harder.
- Start every clip from an approved keyframe. Do not chain several last-frame extensions in a row.
What are the checks before delivery?
At Zero Shot, an AI video studio in London, this is the list we run on character work, whether it is a cartoon series or a brand commercial.
- Make a contact sheet. One frame from each shot, in a grid, beside the reference sheet. Drift that is invisible from one shot to the next is obvious here.
- Check the face. Eye colour, brow shape, nose, jawline, ears, moles, the side of the parting.
- Check age and build. Models quietly make people younger and slimmer.
- Check the wardrobe. Colour, collar, buttons, accessories, which wrist the watch is on.
- Check hands and teeth in every talking shot.
- Check the last frames of each clip. Identity usually slips at the end.
- Watch at full speed on a phone, then frame by frame on a monitor.
- Ask someone who has not seen the project to point at any shot where it looks like a different person.
- Regenerate failures from the keyframe. A wrong face is not something to patch in the edit.
If you have a character that needs to stay the same across a series, an ad campaign or a film, send us a short brief by WhatsApp or email. Tell us how many videos you expect and where they will run, and we will reply with a fixed price.
Questions people ask
Why does my AI character look different in every shot?
Because the model has no memory of the previous shot. Each generation starts again from the prompt, and a description such as 'woman in her thirties with a dark bob' fits millions of faces. Reference images, not longer prompts, are what hold an identity in place.
Which AI video model is best for character consistency?
It changes every few months, so check current documentation. As of September 2026, Veo 3.1, Kling 3.0, Seedance and Runway Gen-4 all accept reference images of a character. The workflow around the model matters more than the choice of model.
Do I need to train a LoRA to get a consistent character?
Not for a single short film. Reference-image features are usually enough for twenty shots. A trained model starts to pay off when the same character has to appear for months, across many videos and several people's hands.
How many reference images does a character need?
A working sheet has a front view, a three-quarter view, a profile, a full-body view and a handful of expressions, all in neutral light. Most tools accept only a few references per generation, so you pick the two or three that suit each shot.