Here is what the problem actually is, and the approach that ended up working.

Why the character drifts

1. Each generation is stateless. The model has no memory of the shot you generated thirty seconds ago. Every scene is a fresh sample from the same distribution.
2. Text is a lossy way to describe a face. "A young girl with brown hair and a red hoodie" is compatible with thousands of faces. Sampling twice gives you two of them.
3. Style and identity are entangled. If you push the prompt to fix the style, you often move the character with it.

The approach that worked

Anchor on an image, not a description. A reference image is a much denser constraint than a sentence. Once the character is pinned to a picture, you stop asking the model to re-invent a face on every scene.

Separate identity from staging. The character reference and the scene description are two different inputs. Scene 1 might be "on a rooftop at sunset", scene 2 "in a kitchen", but the character block stays identical between them.

Lock the look before you animate. Pick one cartoon style for the whole video. Mixing styles mid-video is the fastest way to break the illusion, even if the face holds.

Generate audio after the visuals. Narration and music are laid under a finished timeline. Trying to time visuals to pre-generated audio makes scene boundaries fight the audio.

Where it still breaks

• Consistency holds within one video. Jumping between unrelated projects means a fresh reference.
• Big changes of age or costume mid-video still need a new reference, not just a new prompt.
• Long videos cost more, because every scene is a generation.

I have been building ToonBee AI Cartoon Video Maker, which does exactly this: upload a character image once, then keep that face, outfit and proportions across every scene. It gives you 100 free credits on signup and exports 16:9 or 9:16.

Curious how others are solving character consistency — reference-image conditioning, or something else?