# NovelAI (nai-diffusion V4.5 и V5) — правила промптинга для ИИ-ассистента

> Инструкция для ИИ-ассистента (ChatGPT, Gemini, Grok, Claude, DeepSeek…), который помогает человеку
> писать промпты для NovelAI V4.5 / V5 в ботах AniworldAI. Источник: https://aniworldai.org/engine/nai/agents.md
> Обновлено: 2026-10-17. Сообщество: https://t.me/aniworld_ai · Бот: https://t.me/novelai_bot · https://vk.me/novelai_bot

## Как работать с человеком

- Отвечай на языке человека. Готовый промпт — всегда отдельным блоком кода, чтобы его можно было скопировать целиком.
- Сначала уточни только то, без чего промпт не собрать (кто в кадре, что делает, где, какой стиль). Детальный запрос — не дофантазируй лишнего.
- Правишь чужой промпт — меняй ровно то, о чём попросили; не выкидывай теги молча, предложи и спроси.
- Не выдумывай теги: если не уверен, что тег существует на Danbooru, опиши словами или проверь вики https://danbooru.donmai.us/wiki_pages/.
- Это сервис для арта в духе аниме. Не собирай промпты с откровенным контентом и с персонажами, которые выглядят несовершеннолетними в откровенном контексте.

## Как отправить промпт

Команда `/nai` и промпт: `/nai 1girl, silver hair, rooftop, sunset`. Персонажи — сегментами `!char` (см. ниже). V5 и V4.5 промптятся одинаково: тот же синтаксис тегов, весов и персонажей; у V5 больше бюджет (~1471 токен) и лучше понимание обычных фраз.


---

# Промпты для NovelAI Diffusion V4.5

NAI — **не** Stable Diffusion и **не** ANIMA. Синтаксис свой, привычки ANIMA
сюда переносить нельзя. Ниже — ядро; детали — в справочнике ниже.

## Жёсткие запреты

- ⛔ `(tag:1.3)` — синтаксис SD, в NAI **мёртв**, уезжает мусором в промпт.
  Вес пишется `1.3::tag ::` или скобками `{{tag}}`.
- ⛔ НЕ пиши руками теги качества и эстетики (`masterpiece`, `best quality`,
  `very aesthetic`, `location`, `no text`) — их дописывает в конец сам
  quality-toggle, и он у нас **включён**. Дублировать = жечь бюджет токенов.
  Выключается флагом `!noqt`.
- ⛔ НЕ переноси `(safe:2.0)` из ANIMA — рейтинг тут задаётся иначе.
- ⛔ НЕ дублируй в негативе строки Heavy-пресета: он уже стоит на аккаунте и
  применяется невидимо. Твой негатив — только содержательные добавки.
- ⛔ Никакого Unicode: эмодзи, кандзи, `<3`, `^_^`, символьные danbooru-теги —
  мусор. Только ASCII-английский. (`v` → `peace sign`, `double v` →
  `double peace`, `|_|` → `bar eyes`.)
- ⛔ Экранировать скобки НЕ надо: пиши `shuten douji (fate)`, а не `\(fate\)`.
- ⛔ Не смешивай противоречащие теги — главный источник артефактов:
  `full body` + `close-up`, `sitting` + `standing`, канонический костюм
  персонажа + дописанный `school uniform`.

## Что решает качество (в этом порядке)

1. **Художники** — главный рычаг. Без них V4.5 рисует «дефолтное аниме».
   Боевой формат нашего бота: `1.2::artist:имя::`. Эксперты мешают 3–7
   художников на разных уровнях: ведущий голым или с весом, поддержка
   приглушённая — `{{artist:dino (dinoartforame)}}, artist:s16xue,
   [[artist:fizrotart]]`. Пак несёт не только линию, но и анатомию тел.
2. **Сцена и композиция** — что происходит, поза, свет, кадр.
3. **Гигиена негатива** — анатомия.

Веса — редко и по делу: ~30% топ-работ галереи не имеют ни одного веса.
`{tag}` ×1.05 и стекается, `[tag]` ослабляет, `::` закрывает все скобки.
Уникальный инструмент V4.5 — **отрицательный вес**: `-1::hat ::` снимает
предмет, `-1::monochrome ::` возвращает цвет, `-1::simple background ::,
location` вытаскивает из белой пустоты.

## Негатив

На аккаунте уже включён Heavy-пресет. От себя добавляй только:
- по задаче: `signature, watermark, artist name, artist logo` (подписи),
  `split screen, border` (лишние рамки).

Два закона:
- **Взаимоблок:** `chromatic aberration`, `dithering`, `halftone`, `film grain`,
  `screentone` сидят в Heavy — как эффект в позитиве они не сработают.

## Персонаж

Узнаваемость даёт точный canonical danbooru-тег + тег серии:
`makima (chainsaw man), chainsaw man`. Персонаж надёжно воспроизводится
примерно от 200–300 картинок на danbooru; если их меньше сотни — подпирай
приметами (цвет и длина волос, глаза, фирменная одежда).

Взрослость задаётся автором, а не подразумевается: `aged up` — сильнейший
рычаг, плюс якоря тела (`curvy`/`toned`, `wide hips`, `narrow waist`,
`tall female`) и лица (`mole under eye`, `makeup`). Мёртвые теги, которые
писать бесполезно: `adult`, `older`, `milf`, `mature`, `adult_female`,
голый `tall`. Чтобы не улететь в «за тридцать» — `mature female` и `old`
в негатив.

## Мультиперсонажи — главное преимущество NAI (через бота флагом !char)

```
2girls, onsen, night, steam
!char girl, silver hair, red eyes, towel, source#hug !b3 !neg glasses
!char girl, black hair, blush, target#hug !d3
```
- счёт (`2girls`) — **только** в базовом промпте; сегмент персонажа
  начинается с `girl,`/`boy,` **без числа**;
- `!a1`…`!e5` — место в сетке 5×5: буква = колонка слева направо,
  цифра = ряд сверху вниз (`!c3` — центр). Работают все 25 позиций плюс
  `!auto`, но при двух условиях, и на них люди спотыкаются:
  - **только внутри сегмента `!char`.** Бот режет промпт по `!char` и ищет
    позицию в каждом куске отдельно. В базовой части флаг не разбирается и
    остаётся мусором в тексте:
    ```
    ✅ /n 2girls, on stage !char girl, red hair !b2 !char girl, blonde !e2
    ❌ /n 2girls, red hair !b2, blonde !e2
    ```
  - **только строчными.** `!b2` работает, `!B2` молча игнорируется и уезжает
    в промпт как есть.

  Если человек говорит «бот не понимает твои `!b2` `!e2`» — не спорь и не
  выкидывай позиции, а проверь эти два пункта: почти всегда флаги стоят вне
  `!char`. В `/help` про позиции не написано вовсе, так что человек скорее
  всего просто не знает, как их ставить;
- `!neg` внутри сегмента — персональный негатив этого персонажа, им же
  глушатся протечки между персонажами;
- взаимодействия: `source#hug` у активного, `target#hug` у пассивного,
  `mutual#` у обоих;
- ⚠️ лимит нашего воркера — **5 персонажей**;
- ⚠️ **сегменты персонажей не переводятся с русского** — пиши их сразу
  по-английски, иначе кириллица уедет в NAI как есть;
- character_prompts — норма жанра даже для соло: так сделаны 65 из 100
  работ галереи.

## Флаги бота

`!noqt` (выключить quality-теги) · `!seed N` · `!neg` (заменить негатив) ·
`!addneg` (добавить к пресету) · `!l`/`!large` (1216×832) · `!s`/`!square`
(1024×1024) · `!file` (без пережатия) · `!v`/`!variety` · `!style` ·
`//0.7` в конце — сила img2img.

## Бюджет и текст

Всего ~512 токенов на базовый промпт и всех персонажей вместе — хвост
молча отрезается. Надпись в картинке: `text, english text` в промпте, а
`Text: <слова>` — строго последним, ≤120 символов; quality-toggle с его
`no text` при этом выключать. ⚠️ Через бота базовый промпт приводится к
нижнему регистру, поэтому `Text:` из чата ненадёжен.


---

# Официальный стиль (Safe) на NovelAI V5

## 1. Главный рычаг — канонный иллюстратор тайтла

Пиши `artist:<иллюстратор тайтла>` + тег персонажа + тег тайтла. V5 знает
официальных художников так же хорошо, как веб-художников, и рисует «как в
оригинале»: линию, палитру, пропорции, даже бумагу.

| Тайтл | Художник (тег Danbooru) | Что получилось |
|---|---|---|
| Sousou no Frieren | `artist:abe tsukasa` | обложка манги, поле цветов, закат |
| Spy x Family | `artist:endou tatsuya` | семейный портрет из обложки тома |
| Spice and Wolf | `artist:ayakura juu` | обложка лайт-новеллы |
| Bishoujo Senshi Sailor Moon | `artist:takeuchi naoko` | акварель артбука 1993 |
| Chainsaw Man | `artist:fujimoto tatsuki` | штриховка манги, red theme |
| Genshin Impact | без художника | сплэш-арт Hoyoverse по одним тегам |

Проверяй, что тег художника существует на Danbooru (`/tags.json?search[name]=`).
От ~70 постов уже работает.

## 2. Теги «официальности»

- `official art` (550 тыс. постов) — главный.
- `key visual`, `promotional art`, `cover page`, `novel cover` — жанр кадра.
- `game cg` / `splash art composition` — для гачи: `ultra complexity,
  depthness, dutch angle, from below` + эффекты стихии.
- `anime screencap` — кадр из аниме; с `year XXXX` и `retro artstyle` —
  кадр нужной эпохи (`year 1993`, cel shading).
- `high complexity` — норма; `ultra complexity` — сплэш; `limited palette`
  + `red theme` и т.п. — стилизация манги.
- Медиум из доки NovelAI: `watercolor (medium)`, `traditional media`,
  `painterly`, `ligne claire`, `ukiyo-e`, `art nouveau`.

## 3. Сцена = канонный момент

Официальный арт узнают по детали из канона, а не по позе:
- Фрирен — поле синих цветов и статуя Химмеля;
- Райден — меч из собственной груди (её канонная атака), Глаз Бога в небе;
- Форджеры — газета Лойда, peace sign Ани, капля крови на щеке Йор, Бонд;
- Холо — повозка, пшеница, красное яблоко, монеты;
- Макима — пальцы у рта, кольцевые глаза, жалюзи.

## 4. Мультиперсонаж и текст

- Несколько персонажей: счётчик в базе (`1boy, 2girls`), каждый — в
  `!char boy, …` / `!char girl, …` с тегом персонажа и действием. Работает
  уверенно на 3 персонажа + животное.
- Текст на обложке: `text, english text, the title is written in elegant
  serif letters at the top. Text: Spice and Wolf` — `Text:` строго в конце.
  V5 пишет без ошибок.

## 5. Разница High / Medium на Safe

На Safe-сценах разница между nai5 и nai5m заметно меньше, чем на Q:
стиль иллюстратора держат обе. Medium выдают в основном мелкие детали
(смазанные мелкие предметы, лица на заднем плане).

## 6. Шаблон

```
artist:<иллюстратор тайтла>, <персонаж>, <тайтл>, 1girl, solo, official art,
<key visual | cover page | novel cover | game cg>, high complexity,
<кадр и ракурс>, <канонный момент прозой>, <фирменные предметы>,
<свет и фон>
```

Без `rating:*` и без канонной внешности —
волосы/глаза несёт тег персонажа.


---

# Writing NovelAI Diffusion V4.5 Prompts — A Guide for LLMs and VLMs

You are being asked to write image-generation prompts for **NovelAI Diffusion
V4.5** (an anime/illustration diffusion model, NOT Stable Diffusion, NOT
Midjourney, NOT DALL·E). Its prompt language is specific. This guide gives you
everything to produce prompts that match the quality of expert human users.

It is grounded in three evidence sources, not just documentation:
(a) the official NovelAI docs, (b) ~200 top community works with their full
embedded metadata, (c) tens of thousands of prompts from heavy expert users.
Where the docs and real practice disagree, **real practice wins** and is flagged.

---

## 0. THE ONE THING TO INTERNALIZE FIRST

V4.5 is extremely forgiving of syntax. Prose, tags, typos, underscores, even
"wrong" Stable-Diffusion weights — all produce results. So the quality of an
image is **NOT** decided by punctuation. It is decided by, in order:

1. **Artist tags** (the single biggest lever — see §5).
2. **Scene content and composition** (what's actually happening).
3. **Negative prompt hygiene** (anatomy).

If you obsess over bracket-counting and ignore artists, your output will look
generic. If you nail artists + scene + a clean negative, it will look expert.
Everything below serves those three priorities.

---

## 1. WHAT THE MODEL CAN SEE (hard constraints)

- **Tokenizer: T5.** Budget ≈ **512 tokens total** across the base prompt AND all
  character sub-prompts combined. Overlong prompts silently drop the tail — the
  last things you wrote may never reach the model.
- **No Unicode.** Emoji, kanji, `♥`, `<3`, `^_^`, symbol-tags — all garbage.
  ASCII English only. (Danbooru's symbol tags are renamed: `v`→`peace sign`,
  `double v`→`double peace`, `|_|`→`bar eyes`, etc.)
- **Trained on Danbooru-style tags AND natural language.** You can mix full
  English sentences with comma-separated tags in the same prompt. Invented/rare
  tags often still work ("warm colors", "dramatic lighting").
- Separator between tags: **`, `** (comma + space).

---

## 2. PROMPT ARCHITECTURE

### 2.1 Recommended order
```
[artist tags] → [quality] → 1girl/1boy/count → [character name, series] →
[body/features] → [clothing/state] → [pose/composition/framing] →
[scene/background/lighting] → [effects] → [quality tail]
```
- Put the **most important elements in the first half.** On V3 and older, earlier
  = stronger was a hard rule; on V4.5 it is soft, but "important stuff early"
  still holds, and **style/artist tags belong near the front**.
- Count tags (`1girl`, `2girls`, `6+girls`, `multiple girls`) go early; `solo`
  removes stray background people.
- **Dataset tags must be the very first token** when used: `fur dataset` (furry),
  `background dataset` (peopleless landscapes/still life, photographic).

### 2.2 The "Add Quality Tags" toggle (assume ON)
When ON, the UI silently appends to the END of the prompt. For V4.5 Full:
`, location, very aesthetic, masterpiece, no text`. Consequences:
- Do not hand-duplicate those tags.
- It biases toward "default anime girl" — turn OFF (or bracket `[very aesthetic]`)
  when chasing a specific non-anime art style.
- Its `no text` suppresses SHORT rendered text — turn OFF when doing text (§10).

---

## 3. WEIGHTS (full syntax — but use sparingly)

**Empirical reality: ~30% of top community works use NO weights at all.** Weights
are a tool for resolving *conflicts*, not a seasoning to sprinkle everywhere.

### 3.1 Brace/bracket emphasis (the everyday tool)
- `{tag}` = ×1.05, stacks: `{{tag}}` ≈ 1.10, `{{{tag}}}` ≈ 1.16, and so on.
- `[tag]` = ÷1.05, stacks the same way down.
- Mechanically: each `{` or `]` multiplies everything to its right by 1.05; each
  `}` or `[` divides. A balanced pair localizes the effect. `::` closes all open
  brackets, so `{{{{{rain ::` is legal.

### 3.2 Numeric weights (V4+)
- `1.5::rain, night ::` — weight 1.5 until the closing `::`. Values 0.0–1.0 weaken.
- **`(tag:1.3)` — Stable Diffusion syntax — technically works but is NOT native.**
  Some experts use it out of habit; prefer `1.3::tag ::` or `{{tag}}`.

### 3.3 Negative numeric weights (V4.5 only — no other model has this)
- Remove an object: `-1::hat ::` (stronger: `-3::hat ::`).
- **Invert a concept**: `-1::monochrome ::` (bring color back), `-2.5::flat color ::`
  (add detail), `-1::simple background ::, location` (escape a white void).
- Rule of thumb: **negative weights = targeted removal / inversion; the negative
  prompt = a long list of things to avoid.** Different jobs.

### 3.4 When weights are actually justified
1. Style/artist tags fighting the default anime look (docs themselves stack
   `{{{{style}}}}` 4-5 deep — a lone style tag gets overpowered).
2. A small key detail drowned by a big scene.
3. Negative weights for removal/inversion (§3.3).
4. Softening without deleting (`0.5::coat ::`).
Outside those, plain tags are cleaner.

---

## 4. NEGATIVE PROMPT (Undesired Content)

**Do NOT copy the "minimal negative" habit from other ecosystems.** In NovelAI,
side-by-side comparisons show that with NO undesired content the render is flat,
muddy, sepia; with a proper negative it is clean and sharp. Every expert user
keeps a substantial negative.

### 4.1 Baseline the app already applies (the "Heavy" V4.5 Full preset)
`lowres, artistic error, film grain, scan artifacts, worst quality, bad quality,
jpeg artifacts, very displeasing, chromatic aberration, dithering, halftone,
screentone, multiple views, logo, too many watermarks, negative space, blank page`
(A lighter preset drops the effect-tags. Human-focus adds `@_@, mismatched pupils,
glowing eyes, bad anatomy`.)

### 4.2 What experts add on top (near-universal)
Prepend `nsfw` if you want to steer clear of explicit content, then anatomy:
```
bad anatomy, bad hands, bad feet, extra fingers, fewer digits, extra limbs,
extra arms, missing limbs, deformed, bad proportions
```
Optional, situational: `signature, watermark, artist name, artist logo` (kills
watermarks/artist signatures), `split screen, border` (kills accidental panels/
frames), `text` (kills stray text).

### 4.3 The interlock trap
The Heavy preset already contains `chromatic aberration, dithering, halftone,
film grain, screentone`. If you want any of those AS AN EFFECT in the positive
prompt, they won't render while Heavy is active — use the Light preset or a
negative-weight instead.

### 4.4 Golden rule
POSITIVE prompt, not by banning the adjacent step.

---

## 5. ARTISTS — THE MASTER LEVER

This is what separates expert output from generic output. Every high-quality
community work uses artist tags, in one form or another, usually several at once.

### 5.1 How to write them (all forms seen in the wild, all work)
- Bare name, comma-separated: `incase, rustle, takamichi`
- With prefix: `artist:incase` or `artist: incase`
- Danbooru names use **underscores or spaces interchangeably**: `da_mao_banlangen`
  or `da mao banlangen`. Escape disambiguation parens only if the tag has them:
  `artist:dino (dinoartforame)`.
- **No `@` prefix** (that's an Anima/other-tool convention; harmless but not NAI).

### 5.2 Mixing multiple artists (the expert signature)
Real experts blend **3–7 artists** and balance them with brace levels like a
mixing console — this creates a bespoke style no single artist gives:
```
{{artist:dino (dinoartforame)}}, artist:s16xue, [[artist:fizrotart]], [[[[artist:dorontabi]]]]
```
- `{{...}}` = louder, bare = normal, `[[...]]` / `[[[[...]]]]` = quieter.
- Fractional numeric support also common: `nixeu, 0.45::gogalking, hxd ::` — lead
  artist bare/strong, support artists dialed down.

### 5.3 Practical rules
- **The artist pack carries the render** — skin texture, light, linework, and to a
  large degree body proportions come from it. Choose the pack for the *look* you
  want, not just the character.
- A style that "floats" or looks plastic is usually fixable by adding a strong
  realist/painter artist, not by tweaking tags.
- One weak artist alone rarely holds a style; stack a lead + 2-3 support.
- If you don't know real artist names, you can still get a decent image from
  scene + quality tags alone (some top works use zero artists, pure tag-spam like
  `{{{{naked apron}}}}, {{{wet clothes}}}`), but you leave quality on the table.

---

## 6. TAG VOCABULARY YOU MUST KNOW (Danbooru-native)

The model was trained on Danbooru tags. Use canonical tags, not prose descriptions,
for anything the dataset has a word for. Key families:

- **Framing (narrow→wide):** `close-up, portrait, upper body, cowboy shot,
  feet out of frame, full body, wide shot`. Camera: `from above/below/behind/side,
  from front, profile, straight-on, dutch angle, foreshortening, perspective, pov`.
  Trap: `boots` implies visible feet and pulls the crop down; `close-up` zooms
  into whatever is tagged (torso tags + no face = chest close-up).
- **Body/build:** `petite < skinny`; `curvy < plump < fat < obese`;
  `toned < muscular`; `flat chest < small < medium < large < huge < gigantic
  breasts`. Height also follows canvas aspect (taller canvas → taller figure).
- **Lighting:** `backlighting, rim lighting, sidelighting, dim lighting,
  dappled sunlight, window shadow, candlelight, spotlight, bloom, god rays`.
- **Style/medium:** `traditional media` + `watercolor/oil painting/graphite/ink/
  pastel (medium)`; `impressionism, art nouveau, ukiyo-e, realistic, sketch,
  lineart, painterly, limited palette, monochrome, spot color`. `year 2014` /
  `year 1995` shifts the era of the art style. `realistic` steers away from anime.
- **Color themes** (`red theme`, `blue theme`, …) repaint the whole scene's light
  and air. Combine freely.

Prefer canonical tags; verify a tag actually exists (has post count) rather than
inventing dead ones — but remember rare/invented natural-language phrases often
still work as a supplement.

---

## 7. CHARACTER & CONSISTENCY

### 7.1 Identity = the exact canonical character tag
Recognizability comes from the Danbooru character tag (with disambiguation:
`artoria pendragon (fate)`), NOT from listing traits. Add series tag too.

### 7.2 The 200-300 rule
A character reliably reproduces if they have ~200-300+ Danbooru images. Under ~100,
reproduction degrades and you must prop it up with explicit outfit/feature tags
(hair color/style, eye color, signature garment). Verify tag strength before
leaning on it.

### 7.3 Controlling apparent age (steer toward adult depiction)
Apparent age is set by the author, three levers:
1. **Positive `loli`/`child` tags** force youth.
2. **The artist pack's body prior** (§5.4).
3. **The character's canonical build.**
To keep a depiction clearly adult **in the young-adult band (early 20s), not
middle-aged**, use a two-sided bracket (verified post counts in parentheses):
- **Floor (push older):** `aged up` (36k — the canonical "drawn older than
  source" tag, the single best lever) + body anchors `curvy`/`toned`,
  `wide hips`, `narrow waist`, `tall female` (note: bare `tall` is a dead tag),
  face-maturity `mole under eye` (243k), `makeup`, `lipstick`, and a plain phrase
  `an adult woman in her early twenties`.
- **Ceiling (don't overshoot into 30s/MILF):** put `mature female` AND `old` in
  the **negative** — `mature female` (48k) is real and pulls toward matronly.
Dead tags to avoid (post_count 0): `adult`, `older`, `milf`, `mature`,
`adult_female`, `mature_body`, `sharp_eyes`, bare `tall`.

### 7.4 Multi-character (V4+)
Two ways: (a) the app's per-character sub-prompt boxes, or (b) the `|` separator:
`base prompt | character 1 | character 2`. Put count tags (`2girls`) ONLY in base;
each character sub-prompt starts `girl,` / `boy,` with NO number. Order = placement
(first = left). Interaction verbs take `source#action` (active party) /
`target#action` (passive) / `mutual#action` (both), e.g. `source#hug` + `target#hug`.
Reinforce with a plain-language sentence for reliability.

### 7.5 Multi-character in the aniworld bot (`/nai`) — syntax and position grid
When writing prompts for the aniworld Telegram/VK bot, the separator is `!char`
(aliases: `>char`, `>#`, `--char`, `>|`), one segment per character after the base:

```
2girls, onsen, night, steam
!char girl, silver hair, red eyes, towel, source#hug !b3 !neg glasses
!char girl, black hair, blush, target#hug !d3
```

- **Position flag `!a1`…`!e5`** = a cell on NovelAI's 5×5 placement grid.
  **Letter = column, left→right** (`a`=0.1 … `e`=0.9); **digit = row,
  top→bottom** (`1`=0.1 … `5`=0.9). So `!a1` = top-left, `!c3` = dead center,
  `!b3`/`!d3` = mid-left/mid-right, `!e5` = bottom-right.
  One flag per character, anywhere in the segment (conventionally last); it is
  stripped from the text sent to the model.
- Position flags are parsed ONLY inside character segments. In the base prompt
  they are not recognized and reach the model as junk text — never put them there.
- If ANY character has an explicit flag, coordinate mode (`use_coords`) turns on
  for the whole scene, and characters WITHOUT a flag default to center (`c3`).
  Rule of thumb: if you position one character, give every character an explicit
  cell.
- The grid is a nudge, not a hard constraint — reinforce placement with plain
  words in the base ("boy seated in the center, twin girls flanking him").
- Per-character negative: `!neg <text>` inside the segment (kill attribute
  bleed, e.g. `!neg glasses` on the character who must NOT wear them).
- Up to 5 characters. V4/V4.5 only — on V3 models (anime/furry) character
  segments collapse into the base prompt. The ~512-token budget is SHARED
  across base + all character segments.
- Character segments are NOT auto-translated — write them in English (the base
  prompt does get RU→EN translation).

(Verified working end-to-end since 2026-08-07. Earlier guides/threads may claim
positions are broken or crash the bot; that was true before this date.)

---

## 8. TEXT IN THE IMAGE (V4.5 renders legible English)

Put `text, english text` in the prompt; at the **very end** put `Text: <words>`.
Everything after `Text:` goes into the image, so it must be last. Keep ≤120 chars.
Describe color/placement in plain language ("in yellow text"). Turn OFF quality
tags for short texts (their `no text` suppresses them). Multiple texts separated by
a blank line.

---

## 9. PARAMETERS (what to recommend)

- **Sampler:** `Euler Ancestral` (k_euler_ancestral) or `DPM++ 2M` are the
  defaults; `k_euler` and `ddim_v3` also common in top works.
- **Noise schedule:** `karras` (dominant in ~94% of top works).
- **Steps:** **28** (standard; also 23). More rarely helps.
- **Guidance (scale):** **5–6** standard; community mode is ~5, ranges 4–8. Higher
  = sharper but risks over-saturation (mitigate with the Decrisper /
  `dynamic_thresholding`). SMEA/SMEA-DYN samplers help coherence above 1024².
- **Resolution:** 832×1216 portrait, 1216×832 landscape, 1024×1024 square are the
  free/standard sizes.
- **Seed:** fix it to A/B a single tag change; samplers are non-deterministic so
  reproduction is approximate, not bit-exact.

---

## 10. FOR VLMs: TURNING A REFERENCE IMAGE INTO A PROMPT

When given a reference image and asked to reproduce its *style/composition* (not
copy it), read it in this order and emit tags accordingly:

1. **Medium & style first** → `traditional media, oil painting (medium),
   impressionism` / `realistic` / `anime coloring`, plus a `year 20XX` if the era
   is distinctive. These go near the front.
2. **Framing & camera** → pick the exact framing tag (§6) and angle. Name the crop
   you see (`cowboy shot, from below, foreshortening`).
3. **Subject count & identity** → `1girl` / `2girls`; if it's a known character and
   you're asked to keep them, the character tag; otherwise describe build.
4. **Body & pose** → build tags + a posture tag (`lying, on side, arm support,
   bent over, contrapposto`). One clear pose, no contradictions.
5. **Clothing/exposure state** → name each garment once with its color; use the
   see-through`). Do not tag the same garment twice.
6. **Lighting** → the specific light tag you observe (`backlighting, window shadow,
   dappled sunlight, dim lighting`). Light with a logic of shadows reads as
   craftsmanship; a described light source that casts no shadow reads as fake.
7. **Background** → the setting tags + `indoors`/`outdoors`, plus `blurry
   background, depth of field, bokeh` if the reference has shallow focus.
8. **Effects** → `film grain, chromatic aberration, lens flare` only if present
   AND not blocked by the negative preset (§4.3).

Then: pick an artist pack whose look matches (skin/light/linework), and — crucially
— **do not import tags that change the subject's identity or apparent age from the
reference if they conflict with the request.**
Read the picture, not just its tag list — and never carry a tag
just because it was in the source.

**Beware "metadata ≠ intent":** a reference's embedded tags reflect what its author
did, which may include tricks you should strip (loli tags, watermark-inducing
artists, contradictory framing). Take the style, light, and composition; leave the
author's bad habits.

---

## 11. PROVEN PROMPT STRUCTURES (copy the skeleton, swap the content)

These are real shapes that consistently produce expert-level output.

**A. Artist-pack + scene (the workhorse):**
```
{{{character name}}}, artist:LEAD, artist:SUPPORT1, [[artist:SUPPORT2]],
very aesthetic, best quality, SERIES, SETTING, POSE, EXPOSURE STATE,
LIGHTING, best quality, amazing quality, very aesthetic, absurdres
```

**B. Prose narrative (great for multi-subject / interaction):**
```
LEAD_ARTIST, Character A from SERIES. Character B from SERIES.
{{{A short plain-English sentence describing the key action and who does what.}}}
Setting and lighting described in a sentence. Mood.
```

**C. Realist/photographic:**
```
1girl, {{{character, hair, eyes, build}}}, artist:REALIST1, artist:REALIST2,
realistic, year 2023, {{best quality}}, SETTING, sweat / wet skin / shiny skin,
POSE, from front, close-up, LIGHTING
```

**D. Minimalist (when unsure — let the model breathe):**
13-20 flat tags, no weights: `1girl, CHARACTER, SERIES, STYLE_TAG, build,
state, setting, lighting, quality`. Reliable, low-artifact.

---

## 12. FINAL CHECKLIST BEFORE EMITTING A PROMPT

1. ☐ At least one artist tag (ideally a small pack)? Style near the front?
2. ☐ ASCII only, no emoji/unicode/symbol-tags?
3. ☐ Character = exact canonical tag + series; strong enough (else add anchors)?
4. ☐ Adult depiction anchored, `loli, child, aged down` in negative?
5. ☐ Negative = anatomy tail (`bad anatomy, bad hands, extra limbs…`), plus
   preset already applied?
6. ☐ Any wanted effect NOT sitting inside the Heavy preset?
7. ☐ Each garment / object named exactly once (no "holding a…" duplications)?
8. ☐ One clear pose, no contradictory framing (no full body + close-up)?
9. ☐ Weights only where they resolve a real conflict?
10. ☐ `Text:` (if any) is the very last thing; quality toggle off for short text?
11. ☐ Under ~512 tokens total?

---

## 13. ONE-LINE SUMMARY FOR THE IMPATIENT

*Lead with a small artist pack; write a real scene with clear pose, light, and
one-named garments; keep a proper anatomy negative and adult anchors; use weights
only to resolve conflicts; and when reading a reference, steal its light and
composition but never its author's bad tags.*