This article is published in English.
A Six-Slot Prompt Line to Make Any AI Anime Generator Follow Style
A six-slot prompt structure for anime image generators: subject, pose, outfit, setting, light and style, plus keywords, character consistency, model choice and when to skip it.
Most people start with something like "anime girl with sword, cool, cinematic". The result is a face you've seen hundreds of times, drifting in front of a smear that might be woodland or might be a parking lot. The usual reaction is to switch generators.
The generator is rarely the problem. Guides to anime image generation published this year all point the same way: the prompt carries nearly the whole load, and a typical prompt is one adjective in a disguise. Midjourney's docs say Niji 7 follows details more literally. That sounds like a perk until you realize it also shows up lazy prompts more quickly.
What follows is a structure collected from the most credible guides, trimmed to something you can paste into a prompt box in a couple of minutes. Think of a prompt as a spec sheet, since wishing doesn't work.
Why a bare style word gives you the average
An unfilled slot doesn't stay empty. The model plugs it with the statistical middle of its training data, so vague input yields average output. That explains why "anime girl with sword" gives you the very same girl that about 90,000 other people got. A category was typed in, and no character was described.
The cure is completeness and not length. Forty words that cover six defined slots outperform two hundred words of piled-up adjectives, since each slot settles a question the renderer would otherwise settle for you.
The six-slot line
Put six elements in a fixed order. Then a glance shows what's missing before a credit gets spent.
- Subject. The who or what, with physical traits pinned down. Replace "a girl" with details you could type identically tomorrow, such as a silver bob, green eyes and a small scar on one cheek.
- Pose and framing. The camera's view of the figure: three-quarter, mid-jump, from below, or tight on the eyes. It's the slot skipped most often and the one deciding if the picture feels composed or reads like an ID photo.
- Outfit and props. Name the garment, the fabric and the wear: a navy military coat with frayed cuffs, a sword half out of its sheath. Specifics beat "cool clothes."
- Setting. A wet rooftop, a 1990s Tokyo market street, a tatami room with the sliding door open to a garden. In anime the background holds much of the mood, so an empty setting wastes half the frame.
- Light and colour. Electric-blue neon rim light, warm key from the left against a cool fill, a three-colour palette. Of all six, lighting costs the least effort to improve.
- Style and render. Flat cel shading, hand-inked lines, a shonen key visual, watercolour bleeds. Era and studio vocabulary belong in this slot.
The last slot needs care. A loose style word is how plastic-looking 3D appears when you wanted 2D. Luma's guidance on mecha prompts says plainly that writing "cel-shaded" is what holds off a slide into CGI rendering. That also goes for water, fire, hair and anything else a model would gladly render photoreal if left alone.
Keywords that pull their weight
Some short phrases change a lot.
anime screencap is the strongest retro cue available. It makes the model treat the result as a frame captured from a broadcast and not as a poster, and colour depth and composition both move with it. SpacePrompts noticed this while testing styles, and it combines well with "VHS colour palette" and "subtle film grain."
no characters suits backgrounds. State it and the model devotes its whole effort to the scenery instead of sneaking in a silhouette.
speed lines, diagonal composition separates a figure standing around from a shonen action beat.
Naming a studio can act as shorthand, although some platforms refuse direct references. A phrase like "painterly watercolour animation style" gets close to the soft, warm look with nobody named.
A complete example with every slot filled, meant for editing:
Young courier with a silver undercut and grey eyes, three-quarter view,
crouched on a railing mid-stride; oversized navy delivery jacket with
reflective piping, canvas satchel; rain-soaked Tokyo overpass at 2am,
vending machines glowing behind her; neon magenta rim light, wet asphalt
reflections, muted blue-grey base palette; cel-shaded flat colour,
hand-drawn linework, 90s anime screencap, subtle film grain --ar 3:4
Remove "90s anime screencap" and put in "detailed cinematic key visual", and the era changes while the character stays put. That's what the slots buy you: one thing moves while the others hold.
Keeping the same character in the second image
Product pages leave this out. One eye-catching image is simple in 2026. A matching character in image two is where projects stall: eye colour wanders, hair length grows, the jacket's cut changes. A writer documented it in public with Veo 3. After announcing the consistency problem solved, he reused a prompt that had worked on a second character and got an art style that didn't match at all. The prompt had the same shape and the person was different, so the pair was useless.
Two habits handle most cases.
Produce a character sheet first, and scenes later. Request front, side, back and three-quarter views against white, full body, with the palette shown. Luma's production notes say a sheet used as a reference is how teams keep a character steady through 100+ images. The dull asset comes first.
After that, repeat yourself deliberately. Fixed traits keep identical wording, so a brown trench coat is a "brown trench coat" every time and never "jacket". Keep the camera vocabulary locked as well, and change only pose and expression. A write-up on character design treats this as discipline and not creativity, which is accurate. Every synonym is drift.
Tools that convert a rough idea into a structured prompt help for exactly this reason. Some run every generation through one order: character, environment, lighting, camera. The gain doesn't come from nicer prose. It comes from the template never changing.
Choosing a model per shot
Nobody wins across the board, and whoever awards a single trophy has something to sell. Ropewalk tested one prompt set on four models in late April and divided the result by task. Nano Banana Pro led on cinematic scenes and photo-to-anime conversion, Seedream 4 on shonen character art, and Recraft V4 on clean chibi line work. A roundup from Morphic credits Nano Banana for conversational editing and character consistency, which lines up with the previous problem.
Midjourney's Niji mode still makes the most atmospheric single frames, though recurring characters remain awkward there. Stable Diffusion with Civitai checkpoints offers the most control, in exchange for ComfyUI, LoRAs and a capable GPU.
A sensible routine sends one six-slot line to two or three models and compares the results by eye. That costs pennies: Ropewalk estimates five to thirty cents per generation, whereas a commissioned anime illustration ran $50 to $500 in 2024.
When the structure isn't worth it
Slot discipline pays off for anything that must be reproduced, such as a comic, a cast for a game, or a channel with a recognizable look. For casual play on a lazy Sunday it drains the fun. Write something goofy, press generate, enjoy the result. A half-sentence plus a lucky seed has produced some of the best images anyone has seen.
There's also a limit that never goes away: rights. How generated output can be licensed depends on the platform, and a prompt leaning heavily on one studio's look may leave you somewhere unintended. Read the terms before anything lands on a product page.
A test for today
Write a single character-sheet prompt and keep its wording in a note. Paste it unchanged into three scenes, varying only pose, location and light. A character that survives all three has stopped being a prompt and become an asset you own.