How We Make an AI Animated Series, Step by Step

Every day Natali Production releases a new part of an animated fantasy series. Each one is about a minute long, in vertical 9:16 format, with consistent characters, real dialogue and a cliffhanger at the end. People often ask whether the AI “just makes it”. It does not. Artificial intelligence is the main tool in our workshop, but the series is built by people, step by step. Here is how one episode of Natali and Baudouin or The Mystery Love of Dragon is made.

1. The series bible

Before the first episode there is the bible: a document with the world, its rules and the look of every character. For Natali and Baudouin it describes the elven kingdom of Sylvael, Moonroot Academy, the Heart Tree and its ritual, the orc city of Khazreth and the rule that in this world nobody dies — the defeated are banished “beyond the Veil”. It also fixes small things that viewers notice at once: Natali has warm brown eyes, a thin silver circlet with an opal and two braids with silver leaf charms; Baudouin has a dark beard and pointed elven ears that show through his hair even from behind; every swordsman holds the sword in the right hand.

The bible also contains the season plan. Season 1 of Natali and Baudouin is planned as 60 episodes in six story arcs, so every short part has its place in a bigger story, and every new episode begins with the cliffhanger of the previous one.

2. The screenplay

A one-minute episode has very little room, so the screenplay is strict. It follows a formula we tested on the first parts:

• the strongest image comes in the first two seconds, to stop the scroll;
• then the episode continues the previous cliffhanger;
• one main event and one emotional peak;
• a new cliffhanger in the last five to eight seconds;
• a four-second end card with the logo and the number of the next part.

The dialogue is written in English, short and clear, because every line has to fit into a five-second shot. Then we write subtitles: Russian for every episode, and later translations into 14 languages for YouTube and this site.

3. The storyboard

The screenplay is split into 13–16 shots of about five seconds each. For every shot we decide who is in the frame, where the camera is, what the one smooth action is, and who speaks. A rule that saves a lot of time: one place per scene, one action per shot, faces large. For dialogue we use over-the-shoulder framing: the speaker faces the camera and the listener is seen from behind, so only one character moves their lips.

4. Key frames

For each shot we first create a still key frame with an image model, using approved reference images of the characters and locations. This is where most of the work goes. Every frame is compared with the references: face, nose, eye colour, ears, jewellery, hair, the length of the dress. A model that “forgets” Natali’s eye colour or gives an elf round ears creates a frame that will be rejected. We keep a list of typical mistakes of the image model and check it before writing every new prompt.

Continuity matters as much as likeness. If a sleeve was torn in shot 4, it stays torn in shot 5; if a sword fell on the left, it lies on the left. Before an episode goes further, all first and last frames are put on one contact sheet to check who stands where and what is in the background.

5. Animation

Approved key frames are animated with a video model. Calm scenes — conversations, walks, glances, magic — are rendered on our own computers; scenes with fast action such as jumps, duels and falls are made with a separate model that handles motion better. Each shot gets a detailed prompt: static camera, nobody leaves the frame, the same place, the face stays large. We also give the model the last frame of the shot, so the movement ends where the next shot begins.

The background is part of the acting. When something happens in the academy courtyard, students stand and react — they whisper, point, gasp — and only after the scene do they walk away. We measure background motion in every clip to make sure the world looks alive but does not distract from the heroes.

6. Voices and sound

Every character has a voice chosen once and kept for the whole series, with the same speed and stability settings in every episode. Lines are generated one by one and placed on the timeline under the lip movement. The music is our own soundtrack; background music generated by the video model is not allowed. Sound effects are added only when they match something visible — a blade clinks only when two blades actually touch on screen.

7. Editing and quality control

The episode is assembled, subtitles are burned into a separate version for platforms where viewers watch without sound, and the whole video is checked again: faces over time, jewellery, the corners of the frame, lips moving only with a line, no stray voices, no sudden jumps. If a shot fails, it is made again — but only that shot. Approved clips are never touched again, which keeps the series consistent.

8. Release

A new part is published every day on YouTube, Facebook, Instagram, Threads and TikTok, and on this site with a written recap of the episode. Every video is labelled as AI-generated content. Viewers can watch all episodes for free, with subtitles in 14 languages.

Why we work this way

AI makes it possible for a small studio to release animation every day — something that used to require a large team and months of work per minute. But the tool does not tell a story. The bible, the screenplay, the storyboard, the choice of every frame and every voice, and the decision to throw away a beautiful shot that breaks the character — that is our work, and that is what you see on screen.

Start watching from Part 1 of Natali and Baudouin or Part 1 of The Mystery Love of Dragon, and follow us on YouTube.