Cartoon character creator: designing a character that survives

Why characters drift

A character described only in words is re-imagined on every generation. "A friendly robot" covers millions of possible robots, so the model picks a different one each time — and within a single long clip, a loosely described character can shift as the shot progresses.

This is not a flaw to be prompted around with more adjectives. It is a consequence of how the description works: words define a range, and anything inside that range is a valid answer. To get the same character twice you have to narrow the range until only one answer fits, or stop using words for the job entirely.

Step one: give the character fixed details

Two or three concrete, unusual, physical facts do most of the work. Not "cute" or "friendly" — those are moods, not features. Use things that are either present or absent:

• One blue eye and one green eye
• A dented left shoulder
• A knitted mustard scarf
• Oversized round glasses
• A chipped left ear

Write these the same way every time, word for word. The phrasing itself becomes part of the character's identity as far as the model is concerned, so paraphrasing your own description is a common cause of unexplained drift.

Three details is roughly the sweet spot. One is too loose. Six starts competing for attention with the action you actually want animated.

Step two: lock the design with an image

Once a generation produces a cartoon character you like, stop describing it and start showing it. Capture a clean still from that clip and use it as the opening frame of subsequent videos. Everything visible in that frame — proportions, palette, costume, line weight — carries forward.

This single change removes most consistency complaints. It also changes what your prompt is for: the picture handles appearance, so the text describes only what happens.

Step three: build a reference set for a series

Image input fixes the first frame. Reference input fixes the character, and lets it appear in scenes your reference pictures never showed. That is what you need for a cartoon series, a recurring mascot, or a set of related ads.

A good reference set:

• Two or three angles. A front view plus a three-quarter view resolves most ambiguity about a face. One angle leaves the model guessing about the rest.
• Large and clear in frame. A character lost in a wide landscape gives the model very little to read.
• Neutral lighting. A strong coloured light in the reference gets absorbed into the character's colour scheme and follows it everywhere.
• Plain backgrounds where possible, so the model does not treat scenery as part of the character.
• A consistent short name in the prompt — "the orange cat mascot" — so the words and the pictures agree about who is being referred to.

Designing a character worth keeping

Before any of the technical work, a cartoon character needs a silhouette that reads at thumbnail size. This is the oldest rule in animation and it survives every change in technology: if you cannot recognise the character as a black shape, the design is not working.

Practically, that means one dominant geometric idea — round, tall, boxy, spiky — plus one memorable accessory that breaks the outline. Hats, scarves, ears and tails all do this. Fine detail on the surface does not, because it disappears the moment the video is watched on a phone.

A working order

1. Generate four or five text-only variations to find a design direction.
2. Pick one and write down its fixed details, verbatim, somewhere you can copy from.
3. Capture two or three clean stills of that cartoon character at different angles.
4. Use those as reference images for everything afterwards.
5. Reuse the exact same style clause in every prompt, so lighting and line weight match too.

Once that is done, the character stops being something you re-roll for and becomes an asset you use. Build one in the cartoon video creator, then read the input modes guide for the mechanics of reference generation.