Writing Prompts for Anime-Style AI Art
A good anime prompt is closer to a shot list than to a story. The model does not understand narrative; it understands visual concepts and how strongly you emphasise them. This guide gives you a structure that works across most anime checkpoints and hosted generators, plus a way to improve results methodically.
The basic order
Most models pay more attention to the start of a prompt than the end. Put the important things first:
- Subject: who or what, including that the character is an adult.
- Appearance: hair, eyes, build, clothing.
- Action and pose: standing, seated, looking over shoulder, holding a sword.
- Setting: where they are.
- Lighting and mood.
- Composition and camera: close-up, full body, low angle.
- Style and quality tags: the art style you want and the rendering finish.
You do not need every element every time. Short prompts with five or six strong ideas usually beat long ones where the model has to guess what matters.
Tags versus sentences
Anime checkpoints trained on illustration archives respond well to comma-separated tags:
adult woman, elf ranger, long silver hair, green eyes, leather armor, longbow, forest clearing, morning mist, soft rim light, full body, anime illustration, clean line art
Newer models with stronger text encoders also understand plain sentences:
An adult elf ranger with long silver hair and green eyes, wearing fitted leather armour and holding a longbow, standing in a misty forest clearing at dawn. Clean anime line art with soft rim lighting.
Try both with your tool. Many people end up with a hybrid: one short sentence for the scene, then tags for style.
Always state adulthood
Anime styles can make characters look younger than intended. Counter that explicitly. Words like “adult woman”, “adult man”, “mature”, “in her thirties” or a profession that implies adulthood (“commander”, “office manager”, “tavern owner”) push the model towards adult proportions and faces. Avoid terms that skew young, and never describe childlike features. Reputable generators will refuse prompts that do, and you should not be trying. More on this in safety and age rules.
Example prompts
These are deliberately SFW so you can see the structure. Adjust clothing and mood to taste within your tool’s rules.
Demon queen portrait
adult demon queen, curved black horns, crimson eyes, long dark hair, ornate black and gold gown, seated on obsidian throne, candlelit throne room, dramatic low key lighting, waist-up portrait, anime illustration, detailed line art
Cyberpunk mercenary
adult woman mercenary, short undercut hair dyed teal, cybernetic arm, oversized bomber jacket, rainy neon street at night, reflections on wet asphalt, pink and cyan lighting, medium shot, low angle, modern anime style
Office drama
adult man in his thirties, loosened tie, rolled-up shirt sleeves, tired smile, empty office at night, city lights through window, warm desk lamp, cinematic composition, semi-realistic anime style
Idol on stage
adult idol singer in her twenties, long wavy pastel pink hair, sequined stage outfit, holding microphone, concert stage, spotlights and haze, confetti, dynamic pose, full body, glossy modern anime style
Weighting
Most local tools (and some hosted ones) let you emphasise a word with brackets and a number, for example (silver hair:1.3). Values between 1.1 and 1.4 are usually enough. Going higher tends to distort the image. If a detail keeps being ignored, try moving it earlier in the prompt before you reach for heavy weights.
Negative prompts
A negative prompt lists things you want the model to avoid. Keep it focused on problems you actually see:
lowres, blurry, jpeg artifacts, extra fingers, fused fingers, bad hands, deformed, watermark, text, signature, cropped
Many people also add terms that keep characters clearly adult. A massive copy-pasted negative prompt can flatten your image and fight your positive prompt, so start small and add terms only when a problem appears.
Lighting and composition words that work
- Lighting: rim light, backlight, golden hour, moonlight, neon glow, volumetric light, soft studio light, candlelight, dramatic shadows.
- Camera: close-up, portrait, cowboy shot (mid-thigh up), full body, from above, from below, dutch angle, wide shot, depth of field.
- Mood: melancholic, serene, triumphant, tense, dreamy.
Our colour and lighting guide explains how to combine these.
Iterating with seeds
Random generation is fine for exploration. When you find an image you like, lock its seed and change one thing at a time:
- Copy the seed of the image you like.
- Change one word (hair colour, lighting, angle).
- Generate again with the same seed and compare.
- Keep the change if it helped; revert if it did not.
This turns prompting from guesswork into a controlled experiment. Note that changing the image size, sampler or step count will change the result even with the same seed.
Settings that matter
- CFG scale / guidance: how strictly the model follows your prompt. Around 5–7 is a good start for anime models. High values look harsh and oversaturated.
- Steps: 20–30 is enough for most samplers. More steps rarely fix a bad composition.
- Resolution: generate at the model’s native size (often around 1024 px on the long side for current models), then upscale. Generating very large directly causes duplicated limbs and repeated faces.
Common mistakes
- Stacking ten style names that contradict each other.
- Using names of existing copyrighted characters instead of describing your own design.
- Describing real people. Not acceptable, and blocked by responsible tools.
- Rerolling 200 times instead of fixing one good image with inpainting (see editing and upscaling).
Write less, test more, and change one thing at a time.