Prompting Tips#
How to write a prompt the model you picked actually understands. The right prompt style differs sharply between model families -- a tag list that produces a great Illustrious image produces a bad Flux image, and the reverse is just as true.
Quick reference#
| Model family | Prompt style | Target length | Hard rules |
|---|---|---|---|
| Illustrious / NoobAI | Danbooru tags, comma separated | 20--35 tags | lowercase, spaces not underscores, escape parentheses |
| Pony | Score tag chain + Danbooru tags | 20--35 tags | full score chain, never score_9_up on its own |
| Flux / Flux 2 Klein | Flowing prose | 2--4 sentences | no quality tags, no weight syntax |
| Z-Image / Lumina2 | Descriptive paragraph | 150--250 words | negatives do not apply, no meta-tags |
| Z-Anime | Painterly art-direction prose | 60--280 words | no booru tags, no camera specs unless you want a photo |
| Krea 2 | Expanded descriptive paragraph | one paragraph | quote any text you want rendered |
| Ideogram 4 | Structured caption, built for you | a short idea is enough | defaults to an iPhone-photo look |
| Anima | Natural language sentences | 2--3 sentences | name characters, then describe them |
| Qwen Image / ERNIE / Mage-Flow | Flowing prose | 2--4 sentences | same rules as Flux |
Tip
If you don't want to learn nine styles, turn on AI Enhance. The enhancer carries a hand-written instruction set for each family and rewrites your prompt into that family's house style.
How your prompt is assembled#
On tag-based models, what reaches the image model is not exactly what you typed:
[model prepend] + your prompt + [model append] + [LoRA triggers] + [embeddings]
| Piece | Where it comes from |
|---|---|
| Quality tags | The checkpoint's own configuration, for example masterpiece, best quality, amazing quality. Some models put them before your text, some after. |
| Negative prompt | The checkpoint's configured negative, for example bad quality, worst quality, worst detail, sketch, censor. Your own negative is added on top. |
| LoRA trigger words | Every LoRA you selected contributes its trigger words automatically. |
| Textual-inversion embeddings | lazypos, lazyneg, deep_negative_pony, lazynsfw, smooth_negative -- SDXL-family models only. |
| SFW / NSFW additions | Channel-dependent tags the bot appends. |
The newer families send your prompt through verbatim. Flux, Flux 2, Krea 2, Z-Image, Z-Anime, Ideogram 4, Qwen Image, ERNIE Image and Mage-Flow have no quality tags and no embeddings configured, so what you type (after character matching and enhancement, if you enabled them) is what the model gets. The one exception is NetaYume, which prepends its own instruction preamble.
Do not duplicate the automatic additions
Typing masterpiece, best quality yourself on a model that already adds it produces a duplicate and can degrade the image. The same applies to LoRA trigger words.
The raw switch#
raw: True bypasses all of it -- no prepend, no append, no embeddings, no enhancement. Only your exact text is sent.
/imagine prompt: 1girl, red hair, standing raw: True
raw is also available as a saved setting in /settings.
Illustrious / NoobAI#
Comma-separated Danbooru tags. No sentences.
| Rule | Do | Don't |
|---|---|---|
| Word separator | cherry blossoms |
cherry_blossoms |
| Character parentheses | ganyu \(genshin impact\) |
ganyu (genshin impact) |
| Character plus series | hatsune miku, vocaloid |
miku alone |
| Artist names | wlop |
by wlop |
| Case | all lowercase | Title Case |
| Count | 20--35 tags | 60 tags |
The 20--35 tag ceiling exists because SDXL's text encoder only reads about 75 tokens. Tags past that point are ignored.
Tag order#
- Subject count --
1girl,2boys,solo - Character identity --
hatsune miku, vocaloid - Physical features -- hair, eyes
- Expression
- Clothing
- Pose
- Framing --
upper body,full body,from below - Background
- Lighting
Tags that do nothing here#
score_9 / score_8 / score_7 (those belong to Pony), 8k, 4k, hdr, high resolution, detailed, highly detailed, and masterpiece / best quality (the bot adds those itself).
Before and after#
| You type | What works |
|---|---|
cute anime girl in a cafe |
1girl, solo, smile, long hair, brown hair, brown eyes, blouse, skirt, sitting, cafe, indoors, table, coffee cup, window, warm lighting, looking at viewer, upper body |
hatsune miku singing on stage |
1girl, hatsune miku, vocaloid, solo, aqua hair, twintails, aqua eyes, singing, microphone, stage, spotlight, concert, from below, full body |
epic battle scene with a knight |
1boy, solo, knight, armor, full armor, cape, sword, holding weapon, battle, dynamic pose, battlefield, fire, smoke, dramatic lighting, full body |
Prompt weights#
Weight syntax works on this family: (tag:1.2) strengthens a tag, (tag:0.8) softens it. Default weight is 1.0.
Pony#
Pony reads the same Danbooru vocabulary as Illustrious, but every prompt opens with a fixed chain of score tags.
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up
Warning
Do not use score_9_up on its own. It causes artifacts and broken anatomy. Use the full six-tag chain above.
After the score chain:
| Position | Tag group | Values |
|---|---|---|
| 2 | Source | source_anime, source_cartoon, source_furry, source_pony |
| 3 | Rating | rating_safe, rating_questionable, rating_explicit |
| 4 | Subject | 1girl, solo, species |
| 5 | Features, clothing, pose, setting, style | same vocabulary as Illustrious |
Unlike Illustrious, Pony tolerates natural-language fragments mixed in with the tags.
Before and after#
| You type | What works |
|---|---|
cute pony in a meadow |
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_pony, rating_safe, solo, pony, cute, smile, standing, meadow, grass, flowers, outdoors, sunny day, blue sky |
anime girl with pink hair |
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, rating_safe, 1girl, solo, pink hair, long hair, smiling, looking at viewer, simple background |
Flux and Flux 2 Klein#
Write prose. Two to four sentences describing the scene the way you would describe it to an illustrator.
| Rule | Why |
|---|---|
| No quality tags | masterpiece, best quality, 8k, hdr give no benefit and waste tokens |
| No weight syntax | (word:1.5) is not parsed -- write "with emphasis on the red coat" instead |
| Avoid "white background" | It produces blurry output on Flux dev |
| No vague adjectives | "beautiful", "amazing", "stunning" carry no visual information |
Structure: subject, then appearance, then action, then environment, then lighting, then camera.
Before and after#
| You type | What works |
|---|---|
a wizard casting a spell |
An elderly wizard with a long silver beard and weathered face raises a gnarled wooden staff, blue arcane energy crackling between his fingers. He wears flowing midnight-blue robes embroidered with silver stars, standing in a stone tower chamber lit by candlelight. Medium shot from a low angle, dramatic rim lighting emphasizing the magical glow, mysterious and powerful atmosphere. |
product photo of headphones |
Professional product photography of premium over-ear headphones with matte black finish and brushed aluminum accents, displayed on a smooth marble surface. |
woman in a garden |
A woman in her thirties with auburn hair gathered in a loose braid tends to rose bushes in an English cottage garden. |
The same rules apply to Qwen Image, ERNIE Image, Mage-Flow, and realistic (non-anime) SDXL checkpoints.
Z-Image and Lumina2#
Z-Image reads long, ordinary English through a Qwen3 text encoder.
| Rule | Detail |
|---|---|
| Length | 150--250 words. The sweet spot is 50--180; under 25 words the model has little to work with |
| Negatives | Do not apply. Phrase every constraint positively |
| Meta-tags | 8K, masterpiece, trending on artstation are forbidden -- they degrade the result |
| Metaphor | Avoid it. Describe what is visible, not what it feels like |
| Text in the image | Wrap the exact words in double quotes |
| Shape | One flowing paragraph, not a list |
Long prompts get truncated by the text encoder, which is why the upper limit matters.
Positive phrasing instead of a negative prompt#
| You want to avoid | Write this instead |
|---|---|
| blurry | sharp focus, crisp details |
| bad anatomy | correct human anatomy |
| extra fingers | natural hands with five fingers |
| low quality | high detail rendering |
What to cover#
Subject and quantity, appearance, action, environment, lighting, camera framing, atmosphere. Anything you state explicitly -- named characters, colours, a style you asked for -- is treated as fixed and is not reinterpreted.
Z-Anime#
The Z-Anime fine-tunes want painterly art direction, 60--280 words, in prose. Booru tags such as 1girl, solo, looking_at_viewer actively hurt here, and so do 8K, 4K and trending on artstation. Leave out camera and lens specifications unless you actually want a photographic look.
Four templates cover most requests:
| Genre | Shape of the prompt |
|---|---|
| Character portrait | Detailed anime portrait of [character], soft rim lighting, expressive eyes with detailed reflections, fine hair strands, clean linework, professional anime illustration quality. |
| Action scene | Dynamic anime [scene], dramatic angle, motion energy, speed lines, particle effects, cinematic composition, detailed shading, high quality anime art. |
| Background / landscape | Anime [location] at [time of day], [lighting], [atmosphere], Studio Ghibli inspired detail level, beautiful background art, wallpaper quality. |
| Full scene with characters | A [subject description with traits], shown [pose or action] in [setting]. |
For a known anime or game character, state the series so the canon appearance is anchored.
Krea 2#
Krea 2 wants an expanded, descriptive prompt in one cohesive paragraph.
- Faithfulness first -- everything you stated stays, and detail is added around it, never over it.
- Any text you want rendered goes in quotes, exactly as it should appear.
- Do not over-specify. Pinning down every object in the frame fights the model.
- A prompt that is already detailed only gets light polish.
- Name the medium (photo, oil painting, 3D render) or leave the choice open.
Ideogram 4#
Ideogram 4 does not take a plain sentence. Its "magic prompt" converts your idea into a structured caption describing the whole frame: a high-level description, a style block, and a list of the objects in the scene with their positions. That is why enhancement is always on for Ideogram 4.
What that structure expects from your idea:
| Expectation | What it means for your prompt |
|---|---|
| One coherent subject is one object | Don't split a car into body, wheels and windows -- describe the car |
| Ground, floor, sky, horizon and crowds are background | Describing them as objects is what buries a standing figure's legs in the floor |
| Framing is expressed as how much of the subject is in frame | Say "full body", "three-quarter", or "head and shoulders" |
| Every piece of visible text is its own element | Quote signs, jersey numbers and price labels exactly |
| Names must be named | State the brand, public figure or character -- paraphrase gets dropped |
| No hedging | "things like", "various", "or similar", "oak or walnut" and "implied" phrasing are all rejected |
Ideogram 4 defaults to an iPhone photo look
Unless you ask for something else, photoreal prompts default to a casual iPhone-snapshot aesthetic: off-centre framing, no motion blur, and deliberately no warm grading (the word "warm" is treated as the trigger for the amber "AI look"). If you want cinematic, studio or film photography, say so explicitly. If you want an illustration, say that instead.
Anima#
Two to three descriptive natural-language sentences, standard English capitalisation, no Danbooru tag formatting, and no quality, safety or artist tags. With more than one character, name each character first and then describe them -- that separation is what keeps their features from bleeding together.
General advice#
- Be specific. "warm golden afternoon light through a window" beats "nice lighting" on every family.
- Describe what you want, not what you don't want. Only the SDXL families have a working negative prompt.
- Iterate. Generate once with a short prompt, then refine with 🔄 re-roll and the remix controls.
- Typing a character name can rewrite itself into proper tags -- see Character Matching.
- To restyle an image you already have, use style presets instead of rewriting the prompt.