How to make an AI anime series

Step 1: lock the art style

1. Pick one style and describe it in two lines you will paste everywhere: line weight, shading, color palette, era. "Clean thin linework, cel shading with two tones, soft pastel palette, 1990s TV anime look." Avoid naming living artists or specific studios; describe the qualities instead.

2. Create a style anchor: one full-color illustration of your main character in that style, made with an image model and refined until it is right. Every other reference image is produced to match this one.

In anime, style drift is as visible as face drift; a character drawn with thicker lines in one episode looks like a different show.

Step 2: build model sheets for the cast

3. For each character, make a model sheet the way an animation studio would: front, three-quarter, side and back in the series outfit, plus two or three expressions. Use the style anchor as the input for every image so the style carries over.

Check the details anime viewers notice: hair spikes and their count, eye shape and color, the position of hair clips, uniform badges.

4. Keep the cast small at first. Three to four characters is manageable for a first season. Each one needs its own complete sheet, and each extra character adds review work to every scene they appear in.

5. Size each sheet to the tool's limit. Kling 3.0 Omni accepts up to seven reference images per request, according to Kling's user guide as of October 2026. RefVid takes 1 to 5 images per character, and each character keeps its own set when two or three share a shot. If you must choose, keep each character's front, three-quarter and one expression first.

Step 3: write episodes as shot lists

6. Write a script, then break it into shots of 3 to 10 seconds. A two-minute episode is roughly 15 to 25 shots. For each shot write: which characters appear, the shot size, what happens, and the line of dialogue if any. "Shot 7: Aki, close-up, looks out the train window, then turns toward the camera. Line: we're almost there."

7. Use anime grammar on purpose. Static close-ups with small movement, reaction shots, establishing shots of locations with no characters, and still frames with camera pans are all standard in TV anime. They are also the shots current video models tend to handle best. Save complex action, such as fights and dancing, for a few key moments.

Step 4: generate shot by shot

8. Generate each shot from the references of only the characters in it. Prompts follow the usual scene structure, setting, action, camera, light, plus the two style lines pasted unchanged. Do not describe the characters' looks.

9. Keep two-character shots simple. Two people in one frame is where identities swap or blend. Frame them clearly apart, give each a distinct action, and name each character in the prompt so the model knows who does what; RefVid allows up to 3 subjects in one shot, each named in the prompt. For conversations, use alternating single shots, which is how most anime handles dialogue anyway.

10. Generate establishing shots of each location once and reuse them across episodes. A recurring classroom or street shot gives the series a sense of place and saves generations. In RefVid, an image of the location can also be added to a shot as a scene photo to set the place and look.

Expect to regenerate shots, especially on a first episode.

Step 5: voice, subtitles and editing

11. Assemble in an editor. Lay down the shots in storyboard order, then record or generate voice lines and fit the shots to them. Keep each character's voice settings fixed across episodes. Add subtitles in a clean, readable font; do not ask the video model to draw text on screen.

12. Keep a continuity sheet per episode: which reference version each character used, which locations, any changes. When a character's look needs updating, such as a new outfit for season two, make a new version of their sheet rather than editing the old one.

Common mistakes and checklist

Mistakes: references in slightly different styles; too many characters in season one; crowded multi-character shots; style described differently in each prompt; text and subtitles generated inside the video.

Checklist: two-line style description; style anchor image; model sheet per character; cast of three to four; episodes broken into 3 to 10 second shots; one or two characters per shot; reusable location shots; voice settings fixed; continuity sheet kept.

RefVid saves each cast member once as a character subject, and each character's page lists every clip made from it, which helps when you compare episode one with episode six. Voice, subtitles and editing happen in your own tools.