Why AI Keeps Generating Random Characters (And How to Fix It)

Updated
5 min read

You describe "a young scientist in a lab coat" and get a different person in every scene. You ask for "a modern office" and get a glass tower in one shot and a cramped cubicle in the next. The words were the same each time. The results weren't.

That's because a text prompt describes a category, not a specific person or place. The model fills in everything you didn't say, and fills it in differently every time. Reference images take that guesswork away: instead of describing your scientist, you show the model what she looks like.

What Counts as a Reference in EasyVid

Every character, location, and prop in an EasyVid project has an image. When a scene tags them, as @Name, their images go to the model with that scene:

  • Characters keep the same face, hair, and clothes.
  • Locations keep the same room, street, or landscape, so the furniture doesn't move between shots.
  • Props keep an object the same, like a product, a sword, or a family photo.

How the images are used depends on the workflow. In Classic, they build each scene's image, which can then be animated into a clip. In Director, they go straight to a reference-to-video model with the scene.

How to Add Your Own Reference Images

Attach them when you start. In the Script tab, attach up to five images to your prompt, like a photo of your product, your mascot, or the room you want to use. EasyVid sorts each one into a character, location, or prop, names it, and describes it so the script can tag it.

Upload them later. Open any character, location, or prop and upload a PNG, JPG, or WebP image in place of the generated one. Props are listed under References in the Script tab.

Let EasyVid make them. If you don't have images, EasyVid generates one for every character, location, and prop the script needs. Check them before you generate the scenes. Every scene copies them, so a mistake in one image shows up everywhere.

Change one scene's references. You can add or remove reference images on any single scene and regenerate just that scene.

What Makes a Good Reference Image

The model copies what it sees, including the parts you didn't mean. A few rules help:

  • One subject per image. A character works best as a full-body image on a plain background. A busy photo invites the model to copy the background too.
  • Locations as wide, empty views. Show the whole place without people in it, so characters can be placed into it.
  • Props on their own. Show the object by itself, clearly lit, not in someone's hand.
  • Clear and well lit. Blurry, dark, or tiny images give the model less to hold on to.
  • One look per image. If a character changes clothes, add a variant (@Mara:raincoat) instead of mixing outfits into one image.

Match a Visual Style

Reference images can also set the look of the whole video. In the Script tab's style picker, choose a custom style and upload a style image, such as a painting or a frame from an animated film. EasyVid describes its look and uses that as the project's visual style. If one of your uploaded characters or locations doesn't match that style, EasyVid offers to regenerate it in the style so the cast doesn't look pasted in.

How Many References Fit in a Scene

Each model takes a limited number of reference images per scene, shared by everything tagged in it:

  • Classic scene images: up to 10
  • Director video models: 9 to 11, depending on the model

A scene with three characters, a location, and two props already uses six. If details start slipping in a busy scene, tag fewer things in it. Not every prop needs to be in every shot.

When a Scene Still Doesn't Match

Reference images make results far more consistent, but not pixel-perfect. The model redraws your subject in a new pose, angle, and light, and a small detail can still drift.

One scene is off. Regenerate that scene. The references stay the same, so the next try starts from the same look.

The character is right but in the wrong place. Check that the scene tags the location as well as the character. A scene with no location tag lets the model invent one.

Every scene has the same problem. Fix the reference image itself, then regenerate the scenes that use it.

A character I never mentioned shows up. Look at the scene's tags. Every tagged asset sends its image, so remove the ones that shouldn't be in that shot.

Learn More

The reference images page explains reference-to-video generation and the model limits in more detail. For keeping a cast the same across a whole video, including outfit changes and voices, read how to keep characters the same in every scene.

Ready to create your own AI video?

Transform your ideas into stunning videos with EasyVid's AI-powered tools.

Start Creating

Continue Reading

View all articles