How to make a cartoon with AI: a three-step workflow

Why most AI cartoons turn out badly

A typical beginner’s mistake looks like this: someone opens a video model, writes «a cartoon about a kitten who lost its mother», presses the button and gets six seconds of meaningless motion in which the kitten changes breed twice.

The reason is that a video model is not a director. It will not invent a story, split it into scenes or remember your hero. It does exactly one thing: turns the description of one short shot into a moving picture. Everything else is up to you, before you press generate.

So the workflow does not look like «wrote a prompt, got a cartoon» but like a real animation studio in miniature: script, characters, shoot. It is just that each stage takes minutes instead of months.

Step 1. A script with a second-by-second storyboard

This is the most important step, and it is exactly the one people skip. A script for AI is not a literary text. It is a technical brief that spells out every second: what is in frame, what the character does, where the camera goes, what the sound is.

Writing it by hand takes a long time, so it makes sense to use a text model: ChatGPT, Claude or DeepSeek. They are good at breaking an idea into scenes.

How to brief the text model

Do not ask «write a cartoon script». Ask for a storyboard in the format you need. Something like this:

Break the story into scenes exactly 10 seconds long. For each scene give a table: number, what happens second by second, a shot description for the video generator (composition, light, shot size, camera movement), the character’s lines, background sounds. Write the shot descriptions the way a text-to-video model will understand them: concrete nouns, no abstractions like «an atmosphere of sadness».

Then you edit what it produced. The key point: the shorter the scene, the higher the chance the model will not ruin it. A scene with one action generates well. A scene where the hero stands up, walks, opens a door and looks back will almost certainly fall apart.

The one-action rule

One scene is one action and one camera move. Not «the kitten runs down the street, sees its mother and jumps to her», but three separate scenes. It looks duller in a written script, but that is how a working cartoon is made.

What the shot description must contain

• Shot size: close-up, medium, wide.
• What the character does: one action.
• Camera: static, slow push-in, orbit, sideways track.
• Light and time of day: morning light from a window, night, neon.
• Style: it must be the same in every scene, word for word.

The last point is critical. If in the first scene you wrote «in the style of a Pixar 3D cartoon, soft light, warm palette», that phrase must appear word for word in every other scene. Any rewording is a signal to the model to change something.

Step 2. Character sheets

Now the hero. And here is the main idea of the whole article: a character cannot be described in words.

If in every scene you write «a ginger kitten with a white patch on its chest», you will get ten different kittens. Text describes not a specific hero but a whole class of heroes, and the model picks from that class again each time.

The solution: generate the character once as an image and from then on feed that image as a sample. This technique is covered in detail in the article on character consistency.

How to build a character sheet

Take an image model. Nano Banana Pro (it is strong at editing and keeping the face) or GPT-5 Image (it understands complex descriptions better) work well.

1. Generate the hero face-on. Do not spare attempts: you will see this face a hundred more times. An image costs pennies, and a mistake at this step will ruin the whole cartoon.
2. Make the other angles. Ask the model to show the very same character in profile, three-quarter, full length, in motion. Editing models are exactly the ones that can change the angle while keeping the appearance.
3. Reject. Some shots will be «almost the same» hero. Throw them out: almost means not the same.
4. Assemble the set. That is the character sheet. One for each hero.

Prices at this step are laughable. On KADR an image in Nano Banana Pro costs 16 tokens (about 25 rubles), in Nano Banana 2 eight tokens (13 rubles), in Flux.2 Pro four tokens (6 rubles). A set of five sheets comes to about 125 rubles. Compared with the cost of video that is nothing, so saving here is silly.

Make location sheets too

The same technique works for backgrounds. If the action takes place in a room, generate that room once and then feed it as a reference. Otherwise the furniture will rearrange itself from scene to scene.

Step 3. Animation in Seedance

Now you have a storyboard and heroes. Let us assemble the cartoon.

The Seedance line is the best fit for character animation. There are two reasons. First: it claims character stability between shots and accepts several references at once. Second, purely arithmetic: it generates segments of up to 15 seconds, so a minute of cartoon needs just four generations. Models with an eight-second limit would need seven and a half.

How to generate a scene

In frames mode you feed:

• the character sheets as references (on KADR that is up to 6 images at once),
• the scene description from your storyboard.

And here is the key rule: do not describe the hero’s appearance in the prompt. At all. The model already has the picture. If you give it both the picture and a text description of the face, they will conflict, and the model will produce a compromise, that is, a third character.

The prompt should contain only what the hero does, where they are, how the camera moves and what the style is. The appearance comes from the reference.

Drafts at low resolution

Before paying for final quality, run the scene at 480p. You are checking not the picture but the logic: did the model understand the scene, is the camera going the right way, is the hero falling apart. You do not need resolution for that.

The difference in money is fivefold. The same six-second clip on Seedance 2.0 costs 69 tokens (109 rubles) at 480p and 362 tokens (572 rubles) at 1080p. Ten drafts and one final cost less than three attempts straight at high quality.

Stitching scenes together

So the cartoon does not look like a set of unconnected pieces, use the last-frame technique: take the final frame of a finished scene and feed it as the starting frame of the next. The character, light and setting carry over to the new scene automatically.

A caveat: you cannot stretch this forever, errors accumulate. Every three or four scenes go back to the original character sheets to reset the accumulated drift.

What it really costs

Let us count a minute of cartoon. KADR data, October 2026, 15-second segments. Seedance 2.0 Fast is counted at 720p (it does not go higher), the flagship at 1080p.

This is not a public offer.

Total per minute without regenerations: about 1850 rubles on Seedance 2.0 Fast and about 5850 on the flagship.

Total per minute, honestly: multiply by 3. A cartoon is the hardest genre for AI, because it combines a character, motion and stylistic consistency at once. The real budget for a minute is 5500 to 17 500 rubles depending on the model.

The full calculation for every model is in the article on how much a minute of AI video costs. Sounds like a lot? Compare it with the cost of a minute of classic animation at a studio and with its timelines. And remember that you get the result tonight.

Five mistakes that make a cartoon fall apart

1. No storyboard. You are trying to get a story out of a video model. It will not invent one; it can only do one shot.
2. The character is described in words instead of shown in a picture. You are guaranteed ten different heroes.
3. Several actions in one scene. The model only has time to do one well.
4. The style is reworded from scene to scene. Every rewording is an instruction to change something.
5. Final quality on the first attempt. You pay five times more to test an idea you could have tested at 480p.

If you need an illustrated book: a children’s fairy tale with AI.

In short

A cartoon made with AI is not one button but a three-step pipeline. You write a script with a second-by-second storyboard in a text model, fix the characters as images in Nano Banana or GPT-5 Image, and assemble the animation in Seedance, feeding the hero sheets as references.

You do not need to draw. You need to describe precisely and reject ruthlessly.