How to Make an AI Cartoon Video with Consistent Characters

Getting the same characters to look consistent across multiple AI video shots is the hard part. Anyone can generate a single cute clip. A full episode is different.

This case study walks through a reusable Zoviz Canvas pipeline called Cartoon Mini-Episode Studio: one character sheet, six keyframes generated in sequence, and six 12-second clips that join into a complete 72-second mini episode. The example used is "Pip and the Runaway Scarf," a bedtime short about a little penguin chasing his scarf across a snowy night. Every prompt and the full workflow are included.

Step 0: start with a real screenplay, not a prompt

The biggest mistake people make with an AI video generator is typing one long prompt and hoping a story falls out. This project started the opposite way: with a two page screenplay written first as a normal document, then translated into canvas nodes.

The screenplay defines the logline ("When the wind steals little penguin Pip's beloved scarf, he chases it across the snowy night with his firefly friend Luma, and discovers that the best thing you can do with something you love is give it to someone who needs it more"), three characters, a style bible, and six shots of 12 seconds each, every one with four fields: VISUAL, ACTION, DIALOGUE and SOUND.

Two decisions in that document do most of the heavy lifting later.

A style bible that repeats in every prompt. One fixed sentence, "Soft 3D cartoon style, pastel colors, rounded toy-like shapes, gentle warm lighting, kids TV aesthetic, no text on screen," appears in the character sheet prompt and in all six shot prompts. Repetition is what keeps the rendering style locked.

Shots sized to the video model. Each shot is 12 seconds because the video model behind the canvas (Kling v3 Omni, with audio) generates clips up to 15 seconds. Six shots of 12 seconds gives you a 72 second episode with clean cut points, one clip per shot.

Step 1: the character sheet, your consistency anchor

The first node runs an AI image generator to produce a single 16:9 image of all three characters standing side by side. This is the identity reference every other image in the project points back to.

The prompt behind it:

The canvas README is explicit about what happens next: check this image before generating anything else. If Pip's proportions are wrong here, they will be wrong in all six shots.

Step 2: six keyframes, generated in order

Each shot has its own prompt note feeding an image generation node. The keyframes must run in order, one at a time, because each keyframe uses the previous one as its reference image. That chain, plus the character sheet at its root, is the whole secret of AI character consistency in this workflow.

Every shot prompt follows the same anatomy. Here is shot 1 in full:

Notice the five layers: a shot title, the KEYFRAME description (what the still image shows), the anchor phrase "Same characters as the reference image", the repeated style line, and then VIDEO MOTION and SOUND, which the image node ignores but the video node reads. One prompt drives both nodes.

The six keyframes, in story order:

Same penguin, same firefly, same seal, in six different scenes. That is the payoff of the reference chain.

Step 3: six clips with sound, then assembly

Each keyframe becomes the first frame of a 12 second video clip. The video node reads the VIDEO MOTION and SOUND lines from the shot prompt, so dialogue is spoken and the music is described per shot, from "gentle lullaby music" in shot 1 to "playful chase music" in shot 2 to the full theme returning for the finale. This is an AI video generator with sound built in; no separate voiceover or music step.

The final step happens outside the canvas: download the six clips and join them in order in any editor, CapCut, ffmpeg, whatever you already use. No transitions needed; the shots were written to cut cleanly.

The full run order and what it costs

The canvas ships with a README note pinned to the board, and it doubles as the checklist:

1. Run the character sheet image and check the three characters look right before anything else.
2. Run keyframes 1 to 6 in order, one at a time. Each uses the previous keyframe as reference (about 6 credits each).
3. Check all keyframes, then run the 6 video nodes (12 seconds, about 40 credits each).
4. Download the 6 clips and join them in order in any editor.

Total cost lands around 286 credits per episode: roughly 36 for all the images and 240 for the video. The rule that protects that budget: if you change a shot prompt, rerun that keyframe first and then its video, and always keep the phrase "Same characters as the reference image" plus the style line in every prompt.

The keyframe stage is deliberately cheap compared to the video stage. You iterate where it costs 6 credits, and only render video when the storyboard is exactly right. That's the same discipline a real animation studio applies with animatics before committing to final render.

Why this beats one long prompt

A single "make me a 72 second cartoon about a penguin" prompt fails for three reasons, and each element of this pipeline answers one of them.

Character drift is answered by the reference chain: the character sheet defines the cast, every keyframe points at the previous one, and every prompt re-describes the characters in the same words anyway. Belt and suspenders.

Story structure is answered by the screenplay: six shots with a beginning (the scarf escapes), a chase, a turn (the scarf lands on someone who needs it more), an emotional decision, and a quiet ending. The AI never has to invent structure; it only has to execute one 12 second beat at a time.

Wasted spend is answered by the storyboard economics: bad ideas die at the 6 credit keyframe stage, not the 40 credit video stage.

One more detail worth stealing: shot 6's prompt escalates the no text rule to "ABSOLUTELY NO TEXT: no words, no subtitles, no captions, no letters anywhere in the image or video, dialogue is spoken audio only." Video models love to hallucinate subtitles on emotional closing shots; an explicit hard ban in the final prompt is the fix.

Who this is for

Anything that needs a short animated story with recurring characters: kids channels on YouTube, bedtime story apps, brand mascots for social content, educational micro lessons, church or classroom storytelling, or series pilots you want to test before spending on real animation. Because the canvas is reusable, episode two is a rewrite of six shot prompts, not a rebuild. The character sheet and the structure stay.

FAQ

How do you make an AI cartoon video with the same characters in every scene?

Three habits, all visible in this canvas: generate one character sheet image first and treat it as your identity reference, chain each new keyframe to the previous one as an image reference, and repeat the full character descriptions and style line verbatim in every prompt.

How long can each clip be?

The video model used here (Kling v3 Omni) generates up to 15 seconds per clip with sound. That is why the episode is built from 12 second shots joined in an editor rather than one long generation.

How much does a 72 second episode cost?

About 286 Zoviz credits: roughly 36 in images (one character sheet plus six keyframes at about 6 credits each) and 240 in video (six clips at about 40 credits each).

Does the AI also generate the voices and music?

Yes. Each shot prompt includes VIDEO MOTION and SOUND lines; the clips come back with spoken dialogue, effects and music matching those descriptions. For tighter control you can still replace the audio in your editor.

Can I write the screenplay with AI too?

Yes. Draft it with any LLM, but keep the discipline of this one: a logline, exact character descriptions, a style bible you repeat everywhere, and shots sized to your video model's clip limit.