"Anime" is a broad label covering decades of studios and techniques. To a generative model it is a bundle of visual conventions: cel shading, clean line, large expressive eyes, simplified noses, painterly backgrounds, particular lighting. A prompt should name the conventions rather than the word, because the word alone lets the model pick any of a dozen sub-styles, and it will pick a different one for every scene.

What the model actually knows

Image models have seen millions of anime frames, illustrations and fan art, tagged inconsistently. "Anime" in their training data spans 1980s cel animation, 2000s digital, modern film-grade backgrounds, chibi merchandise art and Western imitations. Naming the conventions narrows the distribution to one corner of that space. Naming an era or a production technique narrows it further; naming a specific studio is unreliable and, for a commercial product, unwise.

Vocabulary that works

Japanese anime style, cel-shaded 2D animation look, clean confident line art, large expressive eyes, soft painterly backgrounds, warm rim light, subtle bloom. Add a camera note (35 mm anamorphic framing) to unify shots, and a colour note (muted teal and amber palette) to unify scenes. Keep the sentence under forty words and put it first in every prompt; models weight the beginning.

For video prompts, add motion vocabulary from the medium: limited animation, held frames, slow camera pan, hair and cloth moving in the wind. Video models asked for "anime" without motion cues produce a photoreal camera move over a drawn frame, which reads as a poster being filmed.

Traps

  • Decorative borders. Image models sometimes draw a frame line or a manga panel edge. Say "fills the whole image edge to edge, no border" and check keyframes before animating; a framed keyframe becomes a framed clip.
  • Realism creep. Cinematic video tiers add photorealistic texture, skin pores and depth of field to flat styles. Fast tiers preserve the drawn look better and cost a fraction. For anime, the cheaper tier is usually the better one.
  • Mixed proportions. Without a reference sheet, one scene draws a character chibi and the next realistic. Sheets fix it; the style sentence alone does not.
  • Humanising the non-human. A firefly becomes a winged girl unless the look says "an actual firefly, an insect" and the sheet shows one. Anime training data is full of humanised animals, and the model reaches for them.
  • Text and captions. Models add Japanese text, sound-effect lettering or a title. Forbid text explicitly; it cannot be removed after animation.

Holding the look across a film

Write the style once as one sentence and prepend it to every cast sheet prompt and every scene prompt. Draw the character sheets first in that style; keyframe the scenes from the sheets; animate the keyframes. Consistency is a pipeline property, not a prompt property: the same sentence in thirty independent text-to-video calls still yields thirty slightly different looks, because each call re-samples. The sheets and keyframes are what pin it.

Backgrounds deserve their own sentence in the style guide, because anime treats them differently from characters: painterly, detailed, static. A keyframe prompt that mentions "painterly background, cel-shaded character" gets the contrast right; one that says only "anime" often produces a cel-shaded background that looks like a game.

Choosing the tier

For anime and other drawn styles, the fast video tier is the right default. It keeps flat colour flat, costs a quarter of the cinematic tier, and its shorter clips (5 seconds) suit limited animation, where holds and cuts are part of the language. Use the cinematic tier only for a scene that needs complex motion, such as water or a crowd, and accept some texture creep there.

Other drawn styles

The same method transfers. Watercolour: wet edges, paper texture, soft bleeding colour, no line. Ink wash: monochrome, brush strokes, negative space. Comic book: bold ink outlines, halftone shading, flat saturated colour, no borders and no speech balloons. Pastel: chalky texture, soft edges, pale palette. In every case: name the medium's conventions, forbid text and frames, fix the sentence, and let sheets and keyframes carry it.

Frequently asked

Why does the model keep adding a caption? Anime frames in training data often carry subtitles or titles. "No text, no caption, no subtitles" at the end of the prompt is usually enough; if not, crop the keyframe before animating.

Can a photo be turned into an anime character? Yes, as a reference for the sheet, with the person's consent. Generate the sheet in the style from the photo, then use only the sheet downstream so realism does not leak.

Do negative prompts work on video models? Inconsistently. Put the forbidden items in the keyframe prompt, where image models honour them, and keep the video prompt about motion only.