Why do my characters look different from scene to scene?
Atlabs keeps a character looking the same across scenes when the video model can use your character as a reference. If a character changes, gets duplicated, or looks different in any scene, it's usually one of the causes below.
Why it happens
Some video models only animate the first image (start frame) of a scene. They can't read character tags (@name). If a scene has no image yet, the model builds the video from the text prompt alone, and the characters won't match. Other models need the References option enabled to retain the existing character context; selecting a model without enabling References can cause characters to change between the image and its motion.
How to fix it
Option A: make the scene image first
- Open the scene and generate the image with your characters.
- Then generate the motion from that image.
Option B: use a reference model
- Open the scene's motion settings and change the model to Seedance 2 Mini Reference Generation.
- Add your characters as reference images.
- Tag them with
@in your prompt, for example@Giggle waves at @Chiku.
If you uploaded your own character
- When you add the character, choose Keep uploaded look.
- Leave the Clothing box empty. Anything typed there can change your character.
- Don't also apply a style conversion to an already-styled image.
Only fix the wrong scenes
You don't need to remake the whole video. Regenerate just the affected scenes, for example, scenes 3, 7, and 17. Regenerating the same image usually fixes duplicates and extra characters.
Example
"I specified a male and a female, but after 2 seconds both look female."
Check the model used for that scene. If it's a start-frame model, generate the scene image with both characters first, then animate it.
Limits
- AI models can still get a few scenes wrong, including changes in character size, duplicated characters, or characters disappearing. Occasional misses are normal. If many scenes are wrong, send us the story link and scene numbers so the team can investigate. Include notes identifying the affected scenes and what is inconsistent (for example, appearance, clothing, or voice). The team can review the issue and follow up; whether credits are refunded depends on the investigation.
- Regenerating costs credits. Fix the image first, because images cost less than video. Regenerate the affected scene image; if it still has errors, try regenerating that scene. You don't need to remake the whole video.
- When several scenes have errors, review and regenerate them scene by scene so you can check each result before generating motion. Character references and prompts can reduce inconsistencies, but AI generation may still produce errors.
Still not right?
Send us your story link and the scene numbers through chat. For widespread inconsistencies, include concise notes about the problems across the video.
Frequently Asked Question
Why do my characters look different once they move?
Motion models may not read character tags. Generate an image with your desired characters tagged first, then generate motion from that image. In the motion prompt, identify each character by visible appearance and position—for example, “the black dog on the table at right drops down while the other puppies laugh.” Make sure the character exists in the story. You can try choosing a different model; check the credits before generating.
Updated on: 10/15/2026
Thank you!
