Your hero has short dark hair and a green jacket in the first scene. By scene four she is blonde, and by scene six she is wearing someone else's coat. Viewers stop following the story, because they can't tell who they're watching.
This guide explains why it happens and how to stop it in EasyVid.
Why AI Characters Drift
Image and video models make each picture or clip on its own. A text description like "a young doctor with dark hair" fits thousands of faces, so every generation picks a new one. Adding more adjectives narrows it down, but never to one person.
The fix is to stop describing the character and start showing them. Give the model the same image of the same character every time they appear, and it draws that person instead of inventing a new one.
How EasyVid Keeps a Character the Same
In EasyVid, every character is an asset with a name, a description, and a portrait. The script tags the character as @Name in each scene they appear in, and the portrait of every tagged character goes to the image or video model with that scene. There is no model or LoRA to train.
When EasyVid writes the script for you, it creates the characters and tags them for you. Your job is to check the cast and the tags.
Step by Step
1. Write or paste the script. Describe an idea, or paste your own script. EasyVid lists every character it finds in the Characters section of the Script tab.
2. Get each portrait right before you generate scenes. The portrait is what every scene copies, so a problem there spreads to every shot. Open each character, read the description, and regenerate the portrait until it looks right. Write clothes and accessories into the description so they stay in every scene. You can also upload your own PNG, JPG, or WebP image instead.
3. Check that every scene tags the character. A scene only gets a character's portrait if the character is tagged. A scene in an EasyVid script looks like this:
[Close-up of @Mara at the stove, stirring a pot, steam in her face @Kitchen]
If you write or edit a scene yourself, use the tag (@Mara), not just the name. A scene that says "Mara" without the @ gets no reference image, and that is where drift creeps back in.
4. Use variants for outfit and state changes. If the story needs a change, like a raincoat, a soaked shirt, or a muddy face, add it as a variant of the same character: @Mara:raincoat. EasyVid makes the variant by editing the original portrait, so the look changes on purpose while the face stays the same. When EasyVid writes the script, it creates the variants a story needs and keeps track of when they start and end.
5. Generate, then fix single scenes. Generate the storyboard and look through it. If one scene drifted, regenerate that scene. The portrait stays the same, so the next try starts from the same look.
Keep the Voice the Same Too
A character who looks right but sounds different in every scene is just as confusing. In EasyVid, each character has one voice for every line they speak, plus a default voice style. Any single line can take a different emotion without changing the voice. In Director, the scene's voice audio goes to the video model, so the character speaks with that voice on screen. See how to make AI voices sound less robotic for directing the delivery.
Common Problems and Fixes
The face is right but the clothes keep changing. Put the outfit in the character's description and make sure the portrait shows it. For a deliberate change, use a variant instead of describing new clothes in the scene.
Two characters blend into one. Give them clearly different descriptions: age, hair, build, and clothing. Similar portraits give the model little to tell them apart.
A character appears in a scene without being mentioned. Check the scene's tags. Remove any tag for someone who shouldn't be in the shot, since each tag sends a portrait.
Crowded scenes lose details. Each scene's characters, location, and props share one set of reference images: up to 10 for Classic scene images, and 9 to 11 for Director's video models, depending on the model. Fewer tagged assets in a shot means each one gets more attention.
My uploaded character looks out of place. If an uploaded image doesn't match the project's visual style, EasyVid offers to regenerate it in that style.
Over-the-shoulder shots show the wrong face. In dialogue, the person with their back to the camera is tagged as a back view (@Mara:back), so the model shows the back of their head, not a second copy of their face.
Keep a Recurring Cast Across Videos
For a series, save your characters to your library and add them to each new project. The same portraits and voices carry over, so episode five matches episode one. Locations and props can be saved the same way.
Learn More
The consistent characters page covers the whole character system. For uploading your own images of characters, places, and objects, read why AI keeps generating random characters or see reference images.