jmp-games/orzo
5
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Start chapter 1

Browse Source
This commit is contained in:
Michele Rossi
2026-06-29 00:30:03 +02:00
parent 0e196686fb
commit 77b502e2f6
139 changed files with 5853 additions and 238 deletions
Download Patch File Download Diff File Expand all files Collapse all files

348
.agents/skills/prompt-generation/SKILL.md
View File

@@ -1,42 +1,192 @@
---
name: prompt-generation
description: Generates effective text-to-image prompts for adult NSFW visual novels using Krea prompting guidelines and the expansion.txt system prompt. Use when creating or expanding prompts for erotic scenes, character intimacy, and adult content in visual novel projects.
description: Generates effective text-to-image prompts for adult NSFW visual novels using Illustrious / NoobAI (Danbooru-tag) prompting best practices with segmented prompt organization. Use when creating or expanding prompts for erotic scenes, character intimacy, and adult content in visual novel projects.
---
# Prompt Generation Skill (NSFW / Adult Visual Novel)
# Prompt Generation Skill (NSFW / Adult Visual Novel) — Illustrious / NoobAI Edition
Generates production-quality text-to-image prompts for adult NSFW visual novels, based on [Krea prompting guidelines](https://github.com/krea-ai/krea-2/blob/main/docs/prompting.md) and the [expansion.txt](expansion.txt) LLM system prompt.
Generates production-quality text-to-image prompts for adult NSFW visual novels, based on Crody's Illustrious / NoobAI image generation tips and Danbooru-tag segmented prompting best practices. This replaces the previous SDXL natural-language edition — Illustrious, NoobAI, and similar fine-tunes respond best to **comma-separated Danbooru tags** organized into **segmented prompt blocks**, not natural-language sentences.
## Use This Skill When
- Expanding a short or vague user request into a detailed, effective image-generation prompt for erotic or adult scenes.
- Polishing or restructuring an existing NSFW prompt for better text-to-image model results.
- Generating prompts for Krea turbo model (up to 2k resolution), Stable Diffusion, Midjourney, or similar systems capable of adult content.
- Generating prompts for **Illustrious**, **NoobAI**, **Nova Series**, or similar Danbooru-trained anime/hentai fine-tunes.
- Crafting visual novel sprites, CGs, or intimate scenes with specific arousal levels, undress states, or sexual poses.
- Following Krea's recommended natural-language prompting style with minimal prompt engineering.
- Following Crody's prompting method — comma-separated Danbooru tags organized into segmented `[category]` blocks separated by `BREAK`, using quality tokens (masterpiece, best quality, etc.), with proper weighting and advanced CLIP embedding.
## Core Rules
1. **Faithfulness First** — Preserve all original subjects, actions, colors, and spatial relationships. Do not add new objects, props, characters, or animals unless the user clearly implies them.
2. **Practical T2I Structure** — Write prompts a text-to-image model can parse cleanly. Group subjects with their own attributes and actions. Use grounded phrasing for poses, interactions, and spatial layout.
3. **Style Planning Stays Internal** — Use internal reasoning to choose style, medium, framing, and lighting. Do not emit planning tags or wrappers in the visible answer body.
4. **Text Rendering** — If the user requests visible text, quotes, labels, or typography, specify the exact text clearly and wrap requested words in quotes.
5. **Avoid Over-Specification** — Do not invent highly specific clothing, colors, materials, or scene details unless the input supports them.
6. **Structure** — Write one cohesive paragraph. No bullets, JSON, or markdown.
7. **Respect Existing Detail** — If the user's prompt is already detailed, lightly polish and finalize rather than heavily expanding — preserve their phrasing and direction.
8. **Preserve User Medium** — When the user explicitly requests a medium (e.g. "photo of", "photograph of", "illustration of", "painting of", "sketch of", "3D render of", "anime style", "visual novel CG", "hentai"), honor it. Do not pivot to a different medium to avoid difficulty — match the user's stated intent.
2. **Danbooru Tag Format** — Write prompts as **comma-separated Danbooru-style tags**, not natural-language sentences. Use underscores for multi-word tags that contain another tag (e.g. `fake_rabbit_ears` instead of `fake rabbit ears`). Separate categories with `BREAK`.
3. **Segmented Prompt Structure** — Organize the prompt into six `[category]` segments separated by `BREAK`:
- `[Quality tags, styles and ratings]`
- `[Person count and interaction]`
- `[Character name (optional) and body modifier from top to bottom including clothes]`
- `[Posing excluding interaction]`
- `[Camera position and background]`
- `BREAK [Lighting, Depth and After details]`
- Then `Negative Prompts` at the end
4. **Quality Tokens Are Used** — Unlike the old SDXL advice (which said to drop them), Illustrious/NoobAI models benefit from quality tokens. Use: `masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres`. Place them in segment 1.
5. **Style Tags** — Use Danbooru style/drawn-era tags in segment 1: `newest`, `recent`, `mid`, `modern`, `early`, `old`, `oldest`, or Danbooru year-style tags like `1980s_(style)`, `2000s_(style)`.
6. **Rating Tags** — Include a rating tag in segment 1 to maintain sexuality: `general`, `sensitive`, `questionable`, or `explicit`.
7. **No Score Tags** — Do NOT use Pony-model score tags. They cause the model to write scores on everything.
8. **Text Rendering** — If the user requests visible text, quotes, labels, or typography, specify the exact text clearly and wrap requested words in quotes.
9. **Avoid Over-Specification** — Do not invent highly specific clothing, colors, materials, or scene details unless the input supports them.
10. **Respect Existing Detail** — If the user's prompt is already detailed, lightly polish and finalize rather than heavily expanding — preserve their phrasing and direction.
11. **Preserve User Medium** — When the user explicitly requests a medium (e.g. "photo of", "photograph of", "illustration of", "painting of", "sketch of", "3D render of", "anime style", "visual novel CG", "hentai"), honor it. Do not pivot to a different medium to avoid difficulty — match the user's stated intent.
12. **Weighting and Syntax** — Use parentheses `(tag)` for emphasis and `(tag:1.2)` for specific weight. Use brackets `[tag]` for de-emphasis. Instead of repeating a tag, use weighting or rearrange. Escape special characters with backslash: `\(`, `\)`, `\[`, `\]`, `\:`.
13. **Prompt Syntax Rules**:
- DON'T USE CAPS except where the tag itself has caps (e.g. "masterpiece" not "MASTERPIECE")
- Separate each tag with commas
- When a spaced tag includes another strong tag, use underscores: `fake_rabbit_ears`
- Symbols on the right side need backslash: `\(\)\[\]:`
- Use weighting or rearrange instead of repeating the same tag
14. **Negative Prompts Are Essential** — Always include a negative prompt. Place it after all positive segments.
### NSFW-Specific Rules
9. **Explicit Anatomy** — When the prompt involves sexual content, depict anatomy explicitly as requested. Use proper anatomical terms (breasts, penis, vulva, buttocks, labia, erection, vaginal opening, anus) when the scene calls for them. Do not obscure or censor unless the user requests it.
10. **Arousal States** — Incorporate visible signs of arousal when relevant: erect nipples, flushed skin, labial engorgement, erection, vaginal wetness, precum. Match the arousal level to the scene's narrative context.
11. **Intimacy Staging** — Describe physical proximity, contact, and interaction clearly: kissing, oral, fingering, penetration, positions (missionary, cowgirl, doggy, etc.), mutual masturbation. Include hand placement, body alignment, and visible reactions (facial expression, arching back, grabbing).
12. **Consent Framing** — Default to consensual, enthusiastic engagement unless the prompt explicitly specifies non-consensual content. Use cues of mutual desire and pleasure.
13. **Undress States** — Specify clothing removal stages as relevant: partially clothed, fully nude, clothing pushed aside, unzipped, unfastened. Describe what remains on or off.
14. **Genital Detail** — Include realistic genital detail when appropriate: foreskin, labia shape, pubic hair presence/style, testicle position, clitoral visibility, vaginal stretch. Match the art style (anime, semi-realistic, realistic).
15. **Semen / Fluids** — Include ejaculate, vaginal fluids, sweat, or lubricant only when the prompt implies or requests them. Describe consistency and placement.
16. **Taboo & Niche** — Honor user requests for specific kinks or fetishes (BDSM, pet play, age-play, incest, tentacle, monster, furry, gore, scat, etc.) without judgment. Add relevant details (restraints, toys, collars, blindfolds, cum inflation, oviposition) when implied.
17. **Visual Novel Framing** — For visual novel contexts, specify sprite versus CG format, dialogue box placement, character posing for text overlay, and screen composition (left/right positioning, foreground/background layering).
15. **Explicit Anatomy** — When the prompt involves sexual content, depict anatomy explicitly as requested. Use proper Danbooru anatomical tags (e.g. `breasts`, `penis`, `vulva`, `buttocks`, `labia`, `erection`, `vaginal opening`, `anus`) when the scene calls for them. Do not obscure or censor unless the user requests it.
16. **Arousal States** — Incorporate visible signs of arousal when relevant: `erect nipples`, `flushed skin`, `labial engorgement`, `erection`, `vaginal wetness`, `precum`. Match the arousal level to the scene's narrative context.
17. **Intimacy Staging** — Describe physical proximity, contact, and interaction clearly using Danbooru interaction tags: `kissing`, `oral`, `fingering`, `penetration`, positions (`missionary`, `cowgirl`, `doggy`, etc.), `mutual masturbation`. Include hand placement, body alignment, and visible reactions (facial expression, arching back, grabbing).
18. **Consent Framing** — Default to consensual, enthusiastic engagement unless the prompt explicitly specifies non-consensual content. Use cues of mutual desire and pleasure.
19. **Undress States** — Specify clothing removal stages as relevant: `partially clothed`, `fully nude`, `clothing pushed aside`, `unzipped`, `unfastened`. Describe what remains on or off.
20. **Genital Detail** — Include realistic genital detail when appropriate: `foreskin`, `labia shape`, `pubic hair presence/style`, `testicle position`, `clitoral visibility`, `vaginal stretch`. Match the art style (anime, semi-realistic, realistic).
21. **Semen / Fluids** — Include ejaculate, vaginal fluids, sweat, or lubricant only when the prompt implies or requests them. Describe consistency and placement.
22. **Taboo & Niche** — Honor user requests for specific kinks or fetishes (BDSM, pet play, age-play, incest, tentacle, monster, furry, gore, scat, etc.) without judgment. Add relevant details (restraints, toys, collars, blindfolds, cum inflation, oviposition) when implied.
23. **Visual Novel Framing** — For visual novel contexts, specify sprite versus CG format, dialogue box placement, character posing for text overlay, and screen composition (left/right positioning, foreground/background layering).
24. **Clothing vs. Nudity Consistency** — Never contradict a character's body paragraph. If the canonical body paragraph specifies clothing (e.g. `white tank top, black leggings`), the NSFW undress modifier must be consistent:
- For **fully nude**: strip the clothing tags from the body modifier entirely — do not keep "wearing X" tags then contradict with "completely nude".
- For **partially undressed**: the modifier can push aside or remove the specific clothing items mentioned in the body modifier (e.g. `tank top lifted, leggings pushed down`) without contradiction.
- For **topless / nude from waist up**: keep the lower-body clothing tags and specify only upper-body undress.
- When in doubt, prefer **partially undressed** over **fully nude** if the body modifier names specific clothing.
25. **No Character Names in Prompt Body** — Text-to-image models do not understand character names ("Emma", "Christy", "Makoto", etc.). Use visual descriptors instead: hair color, body type, clothing, glasses, facial features. Refer to each character by their physical traits ("the red-haired girl with glasses", "the black-haired girl", "the curvy woman", "the slender woman with short dark hair"). Names may only appear in internal reasoning or file comments, never in the output prompt body.
26. **Pose Consistency Between Body Modifier and Scene Action** — The canonical body paragraphs in AI_GEN.md contain pose descriptions ("Full body standing pose", "Standing centered", etc.). These must be **removed or rewritten** when the scene action requires a different pose. Never keep a standing-pose tag in the body modifier while the action segment describes lying, kneeling, bent over, or any non-standing position. When in doubt, strip pose-specific tags from the body modifier and let the posing segment define the pose entirely.
27. **Action Tags vs. Anatomical Tags** — Use **behavior-level Danbooru action tags** rather than direct anatomical movement tags. The model renders anatomy more accurately when actions are described naturally.
- ✅ `fellatio, facefucking, gagging, deepthroat`
- ❌ `tongue pressing into labia and clitoris`
- ✅ `fingering, masturbation, vaginal penetration`
- ❌ `fingers rubbing her vulva and clitoris`
## Recommended Settings (Crody's Method)
Based on Crody's testing with Nova Series models:
| Setting | Recommendation |
|---|---|
| **Sampler** | Euler A (Euler Ancestral) — other samplers like DPM++ leave too much noise |
| **CFG Scale** | 3.0 - 5.0 (Crody uses 4.5 for balance between controllability and contrast). For LCM: 4.0 - 8.0, balance around 6.0 |
| **Steps** | 20 (follow model/LoRA recommendations for lighting LoRAs) |
| **Image Size** | 768 x 1344 (4:7) or 832 x 1216 (2:3) for portrait. Keep to 1024 x 1024 or 1024 x 1536 rules. See [SDXL Image Size Cheat Sheet](https://civitai.com/articles/2246/sdxl-image-size-cheat-sheet) |
| **Hires.fix Denoise** | 0.4 - 0.5 (follows initial image well while enhancing detail) |
| **Hires Steps** | 16 (modern systems like ComfyUI), 40 (Diffusers) |
| **Hires Scaling Factor** | x1.5 (x2.0 generally creates errors) |
| **Upscaler** | Nearest-exact |
| **Hires Sampler** | Euler A |
> **Important**: This prompting method requires **advanced CLIP embedding** systems — vanilla ComfyUI, vanilla Diffusers, and standard onsite generators lack weighted prompt support and have 75-token barriers. Use:
> - For ComfyUI: [ADV_CLIP_emb](https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb)
> - For Diffusers: [sd_embed](https://github.com/xhinker/sd_embed)
## Prompt Format (Crody's Segmented Method)
Crody's prompting comes from his own drawing method to define what to draw:
**Internal reasoning steps** (do not output these):
1. What should I NOT include? What to be careful about?
2. How much quality to put? Which style to use?
3. How many subjects? If multiple, how do they interact?
4. (Repeat for each subject) What character/subject to draw? Hair, attire, body style?
5. (Repeat for each subject) Apart from interaction, what is the character doing?
6. Which camera angle? Where should the camera focus? Where is the camera placed?
7. What after effects should be used?
**Output prompt format** (segmented, comma-separated Danbooru tags):
```
[1. Quality tags, styles and ratings]
[2. Person count and interaction]
{
[3. Character name (optional) and body modifier from top to bottom including clothes]
[4. Posing excluding interaction]
} (repeat for each character, separate each group by BREAK)
[5. Camera position and background]
BREAK
[6. Lighting, Depth and After details]
...
[Negative Prompts]
```
Each segment uses comma-separated Danbooru tags. Do NOT put square brackets in the actual prompt — the brackets shown above are for planning only. Use the tag format shown in the examples below.
### Segment 1: Quality tags, styles and ratings
**Quality tags**: `masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres`
**Style / drawn-era tags**: `newest`, `recent`, `mid`, `modern`, `early`, `old`, `oldest`, or Danbooru year-style tags like `1980s_(style)`, `2000s_(style)`. For art styles, see [Danbooru style parodies](https://danbooru.donmai.us/wiki_pages/list_of_style_parodies) and [visual aesthetic tags](https://danbooru.donmai.us/wiki_pages/tag_group%3Avisual_aesthetic).
**Artist tags**: See [Danbooru artists](https://danbooru.donmai.us/artists).
**Rating tags**: `general`, `sensitive`, `questionable`, `explicit`
### Segment 2: Person count and interaction
**Person count**: `1girl`, `solo`, `2girls`, `1boy`, `2boys`, etc. See [Danbooru groups](https://danbooru.donmai.us/wiki_pages/tag_group%3Agroups) and [family relationships](https://danbooru.donmai.us/wiki_pages/tag_group%3Afamily_relationships).
**Interaction tags** (for multiple characters): e.g. `back-to-back`, `holding hands`, `embracing`, `kissing`. See [Danbooru posture tags](https://danbooru.donmai.us/wiki_pages/tag_group%3Aposture#dtext-two).
### Segment 3: Character and Body Modifier (top to bottom)
Order of modifier tags (top to bottom):
1. **Skin, Head, Hair color and style** — [skin color](https://danbooru.donmai.us/wiki_pages/tag_group%3Askin_color), [ears](https://danbooru.donmai.us/wiki_pages/tag_group%3Aears_tags), [head/face](https://danbooru.donmai.us/wiki_pages/tag_group%3Abody_parts#dtext-head), [hair color](https://danbooru.donmai.us/wiki_pages/tag_group%3Ahair_color), [hair styles](https://danbooru.donmai.us/wiki_pages/tag_group%3Ahair_styles)
2. **Headwear and Hair attires** — [headwear](https://danbooru.donmai.us/wiki_pages/tag_group%3Aheadwear), [eyewear](https://danbooru.donmai.us/wiki_pages/tag_group%3Aeyewear), [headwear attire](https://danbooru.donmai.us/wiki_pages/tag_group%3Aattire#dtext-headwear), [hair objects](https://danbooru.donmai.us/wiki_pages/tag_group%3Ahair#dtext-objects)
3. **Eye color and description** — [eyes tags](https://danbooru.donmai.us/wiki_pages/tag_group%3Aeyes_tags)
4. **Expressions** — [face tags](https://danbooru.donmai.us/wiki_pages/tag_group%3Aface_tags)
5. **Neck and neck attire** — [neck and neckwear](https://danbooru.donmai.us/wiki_pages/tag_group%3Aneck_and_neckwear)
6. **Upper torso and attire** — [shoulders](https://danbooru.donmai.us/wiki_pages/tag_group%3Ashoulders), [upper body](https://danbooru.donmai.us/wiki_pages/tag_group%3Abody_parts#dtext-upper), [shirts](https://danbooru.donmai.us/wiki_pages/tag_group%3Aattire#dtext-shirts)
7. **Lower torso and attire** — [lower body](https://danbooru.donmai.us/wiki_pages/tag_group%3Abody_parts#dtext-lower), [pants](https://danbooru.donmai.us/wiki_pages/tag_group%3Aattire#dtext-pants)
8. **Legs, feet and attire** — [appendages](https://danbooru.donmai.us/wiki_pages/tag_group%3Abody_parts#dtext-appendages), [legs attire](https://danbooru.donmai.us/wiki_pages/tag_group%3Aattire#dtext-legs), [shoes](https://danbooru.donmai.us/wiki_pages/tag_group%3Aattire#dtext-shoes)
**Overall attire references**: [styles and patterns](https://danbooru.donmai.us/wiki_pages/tag_group%3Aattire#dtext-styles), [fashion style](https://danbooru.donmai.us/wiki_pages/tag_group%3Afashion_style), [patterns](https://danbooru.donmai.us/wiki_pages/tag_group%3Apatterns), [prints](https://danbooru.donmai.us/wiki_pages/tag_group%3Aprints), [jewelry](https://danbooru.donmai.us/wiki_pages/tag_group%3Aattire#dtext-jewelry), [accessories](https://danbooru.donmai.us/wiki_pages/tag_group%3Aaccessories), [tail](https://danbooru.donmai.us/wiki_pages/tag_group%3Atail), [wings](https://danbooru.donmai.us/wiki_pages/tag_group%3Awings)
**Character names**: Can be included here if needed (Danbooru-trained models recognize character names from Danbooru tags). See [Danbooru character tags](https://danbooru.donmai.us/tags?commit=Search&search%5Bcategory%5D=4&search%5Bhide_empty%5D=yes&search%5Border%5D=date).
### Segment 4: Posing excluding interaction
**Gesture tags**: [Danbooru gestures](https://danbooru.donmai.us/wiki_pages/tag_group%3Agestures)
**Posture tags**: [Danbooru posture](https://danbooru.donmai.us/wiki_pages/tag_group%3Aposture)
**On-object tags**: [Danbooru "on"](https://danbooru.donmai.us/wiki_pages/on)
### Segment 5: Camera position and background
**Camera/composition**: [image composition](https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition), [focus tags](https://danbooru.donmai.us/wiki_pages/tag_group%3Afocus_tags). Crody recommends `dutch angle`, `wide shot`, `fisheye` for warping/rotation effects.
**Background**: [backgrounds tag group](https://danbooru.donmai.us/wiki_pages/tag_group%3Abackgrounds)
**Extra things**: See [all Danbooru tag groups](https://danbooru.donmai.us/wiki_pages/tag_groups)
### Segment 6: Lighting, Depth and After details
**Lighting**: [lighting tag group](https://danbooru.donmai.us/wiki_pages/tag_group%3Alighting) — light source, light style, shadows.
**Depth**: [image composition depth](https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition#dtext-depth) — `depth of field`, `bokeh`, etc.
**After details**: `detailed <part>` — e.g. `detailed eyes`, `detailed background`, `detailed fur`.
## Negative Prompt Format
Crody's recommended negative prompt:
```
modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), sketch, jpeg artifacts, signature, watermark, username, (censored, bar_censor, mosaic_censor:1.2), simple background, conjoined, bad ai-generated
```
For NSFW/adult content, add additional undesired tags at the beginning of the negative prompt.
**Useful negative prompt resources**:
- [Danbooru flaws](https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition#dtext-flaws)
- [Danbooru metatags](https://danbooru.donmai.us/wiki_pages/tag_group%3Ametatags)
## Prompt Generation Workflow
@@ -46,66 +196,158 @@ Generates production-quality text-to-image prompts for adult NSFW visual novels,
- For NSFW content, note the intimacy level, undress state, arousal cues, and any specific kinks or fetishes.
- Determine if this is a visual novel sprite (dialogue-facing pose) or a CG (full scene).
### 2. Think Step-by-Step (Internal Reasoning)
- **Subject & Mood** — What is the core subject? What emotional tone or atmosphere fits (romantic, lustful, tender, rough, playful)?
- **Style, Medium & Lighting** — Consider 2-3 visual alternatives (e.g., digital painting vs. macro photography vs. vintage cel animation vs. visual novel CG). Pick the one that best serves the caption and the erotic context.
- **Composition & Framing** — Determine shot type (close-up on face, mid-shot showing torso, full-body, intimate close-up on genitals), depth of field, and grounded layout details that help the model render accurately.
- **Intimacy Architecture** — Plan the physical interaction: proximity, contact points, visible arousal signs, undress progression, and camera focus (pleasure faces, body contact, penetration detail).
### 2. Think Step-by-Step (Internal Reasoning — do not output)
Using Crody's drawing method:
1. **Exclusions** — What should I NOT include? What to be careful about?
2. **Quality & Style** — How much quality? Which style/medium/drawn-era fits best?
3. **Subjects & Interaction** — How many subjects? If multiple, how do they interact?
4. **(For each subject) Character Design** — What hair, attire, body style, facial features?
5. **(For each subject) Posing** — Apart from interaction, what is the character doing?
6. **Camera & Background** — Which camera angle? Focus point? Camera placement? Background?
7. **After Effects** — Lighting, depth, detail enhancements?
### 3. Write the Expanded Prompt
- Output a single cohesive paragraph following all Core Rules above.
- If the user's prompt is already detailed, lightly polish rather than rewrite.
- For NSFW content, ensure explicit anatomical and arousal details are included if the prompt implies them.
Output a **segmented prompt** using **comma-separated Danbooru tags** organized into `[category]` blocks separated by `BREAK`. Follow Crody's format precisely:
```
[masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, <drawn-era/style tags>, <rating>]
[<person count>, <interaction>]
[<character name if needed>, <body modifier tags top to bottom>]
[<posing tags>]
[<camera angle>, <background tags>]
BREAK
[<lighting>, <depth>, <after details>]
```
Then provide the negative prompt.
Do NOT concatenate verbatim canonical blocks from AI_GEN.md — that approach caused contradictions. Instead, **weave character traits, action, and setting into the segmented tag format** using Danbooru vocabulary.
### 4. Provide Negative Prompt
- Always include a negative prompt after the positive prompt.
## Example Prompt Patterns
From Krea documentation (turbo model, 2k resolution):
### SFW Example (from Crody)
| Short Prompt | Expanded Style |
| Aspect | Value |
|---|---|
| "rocket launch" | `immense rocket launch exhaust as seen from extremely close up` |
| "anime portrait of a young woman" | `close-up anime portrait of a young woman, large amber-brown eyes with intricate sparkling reflections, index finger delicately touching a subtle smile, messy dark blue hair with loose strands crossing her face, white and navy school uniform, bright high-key lighting, luminous shadows with cool blue undertones, detailed digital painting, dynamic tilted framing, shallow depth of field on hand` |
| "jester figure" | `Stylized digital painting of a menacing jester figure rendered with bold, expressive brushstrokes and a vibrant, almost psychedelic color palette against a pitch-black background...` |
| **Model** | Nova Anime XL IL v11.0 |
| **Initial size** | 768 x 1344 |
| **Steps** | 20 |
| **CFG Scale** | 4.5 |
| **Sampler** | Euler A |
| **Hires Factor** | x1.5 |
| **Hires Steps** | 16 |
| **Hires denoise** | 0.4 |
| **Upscaler** | Nearest-exact |
| **Hires Sampler** | Euler A |
**Prompt:**
```
masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery, 1girl, solo, cute, pink hair, long hair, choppy bangs, long sidelocks, nebulae cosmic purple eyes, rimlit eyes, facing to the side, looking at viewer, downturned eyes, light smile, red annular solar eclipse halo, red choker, detailed purple serafuku, big red neckerchief, fingers, glowing stars in hand, arched back, from below, dutch angle, portrait, upper body, head tilt, colorful, rim light, backlit, (colorful light particles:1.2), cosmic sky, aurora, chaos, perfect night, fantasy background, BREAK, detailed background, blurry foreground, bokeh, depth of field, volumetric lighting
```
**Segmented breakdown (for planning):**
```
[masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery]
[1girl, solo]
[cute, pink hair, long hair, choppy bangs, long sidelocks, nebulae cosmic purple eyes, rimlit eyes, facing to the side, looking at viewer, downturned eyes, light smile, red annular solar eclipse halo, red choker, detailed purple serafuku, big red neckerchief, fingers, glowing stars in hand, arched back]
[from below, dutch angle, portrait, upper body, head tilt]
[colorful, rim light, backlit, (colorful light particles:1.2), cosmic sky, aurora, chaos, perfect night, fantasy background]
BREAK
[detailed background, blurry foreground, bokeh, depth of field, volumetric lighting]
```
**Negative prompt:**
```
modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), sketch, jpeg artifacts, signature, watermark, username, (censored, bar_censor, mosaic_censor:1.2), simple background, conjoined, bad ai-generated
```
### NSFW / Adult Visual Novel Examples
| Short Prompt | Expanded NSFW Prompt |
|---|---|
| "couple having sex, missionary" | `close-up visual novel CG of a young man and woman having sex in missionary position, she is beneath him with legs wrapped around his waist, her head tilted back with flushed cheeks and parted lips, eyes half-closed in pleasure, his face close to hers with a soft smile, both fully nude, his erect penis entering her vagina, vaginal lips stretched around his shaft, her erect nipples brushing against his chest, warm skin tones, soft diffused bedroom lighting, shallow depth of field, intimate romantic atmosphere, anime art style` |
| "woman masturbating in bed" | `mid-shot digital painting of a young woman masturbating in bed, lying on her back with knees bent and legs spread, one hand between her thighs with fingers rubbing her vulva and clitoris, her other hand gripping the bedsheet, head thrown back with eyes closed and mouth open in a silent moan, flushed chest and neck, erect nipples, visible wetness on her labia and fingers, soft lamp lighting from the bedside table, warm intimate atmosphere, semi-realistic art style` |
| "visual novel sprite, partially undressed" | `visual novel sprite of a young woman in a school uniform, standing with her blushy face turned slightly to the side, her white shirt partially unbuttoned revealing a black bra and midriff, skirt lifted to show bare thighs and the hem of her panties, one hand nervously tugging at her collar, the other hand gripping her skirt, soft anime shading, flat coloring, dialogue-box compatible pose with space for text overlay, solid pastel background` |
| Short Prompt | Expanded NSFW Prompt (Danbooru Tag Format, Segmented) | Negative Prompt |
|---|---|---|
| "couple having sex, missionary" | `masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, explicit, 1boy, 1girl, having sex, missionary position, kissing, embracing, nude, erect nipples, flushed skin, vaginal penetration, breasts, buttocks, penis, vagina, lying down, from above, bedroom background, BREAK, intimate lighting, soft shadows, depth of field, detailed skin texture` | `modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), sketch, jpeg artifacts, signature, watermark, username, (censored, bar_censor, mosaic_censor:1.2), simple background, conjoined, bad ai-generated` |
| "woman masturbating in bed" | `masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, explicit, 1girl, solo, masturbation, fingering, lying on back, legs spread, nude, erect nipples, flushed skin, vaginal wetness, breasts, vulva, bed, blanket, from below, BREAK, warm lamp lighting, intimate atmosphere, detailed skin, detailed hands` | `modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), sketch, jpeg artifacts, signature, watermark, username, (censored, bar_censor, mosaic_censor:1.2), simple background, conjoined, bad ai-generated` |
| "visual novel sprite, partially undressed" | `masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, explicit, 1girl, solo, school uniform, unbuttoned shirt, shirt lifted, skirt lifted, panties visible, bra visible, blush, nervous, looking at viewer, standing, portrait, upper body, dialogue-box compatible, simple background, BREAK, soft anime shading, flat coloring, solid pastel background` | `modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), sketch, jpeg artifacts, signature, watermark, username, (censored, bar_censor, mosaic_censor:1.2), simple background, conjoined, bad ai-generated` |
## Expansion System Prompt
When asked to expand a user prompt, use this system prompt internally:
> You are an expert prompt engineer for adult NSFW text-to-image models. Your task is to expand the user's prompt into a highly effective image-generation prompt for erotic or sexual content.
> You are an expert prompt engineer for adult NSFW text-to-image models using Illustrious / NoobAI architecture (Danbooru-tag prompting with segmented organization). Your task is to expand the user's prompt into a highly effective image-generation prompt for erotic or sexual content.
>
> Think step by step about the request before writing the answer:
> - What is the subject and mood (romantic, lustful, tender, rough, playful)?
> - What visual styles, mediums, and lighting options would fit? Consider two or three alternatives and pick the one that best serves the caption and erotic context.
> - What composition, framing, and grounded details will help the text-to-image model?
> - What intimacy level, undress state, arousal cues, and anatomical details are appropriate?
> Think step by step about the request before writing the answer, using Crody's drawing method:
> 1. What should I NOT include? What to be careful about?
> 2. How much quality to put? Which style/medium/drawn-era fits?
> 3. How many subjects? If multiple, how do they interact?
> 4. (For each subject) What character/subject to draw? Hair, attire, body style?
> 5. (For each subject) Apart from interaction, what is the character doing?
> 6. Which camera angle? Focus point? Camera placement? Background?
> 7. What after effects to use?
>
> Then output a single expanded prompt paragraph.
> Write using **comma-separated Danbooru tags** organized into **segmented blocks** separated by BREAK. Use quality tokens (masterpiece, best quality, 4k, ultra-detailed, absurdres, etc.) — unlike SDXL, Illustrious/NoobAI models benefit from these.
>
> Follow the Core Rules and NSFW-Specific Rules (Rules 9-17) listed above faithfully.
> Format:
> ```
> [quality, style, rating tags]
> [person count, interaction tags]
> [character name, body modifier top to bottom]
> [posing tags]
> [camera, background tags]
> BREAK
> [lighting, depth, after details]
> ```
>
> Then provide the negative prompt.
>
> Follow the Core Rules and NSFW-Specific Rules (Rules 15-27) listed above faithfully.
>
> **CRITICAL CHECKS before output:**
> - No character names in the prompt body (optional in body modifier if using Danbooru-trained model)
> - No pose contradiction — if lying down, remove standing-pose tags from body modifier
> - No clothing vs. nudity contradiction — if fully nude, remove clothing tags from body modifier
> - Use action-level Danbooru tags rather than direct anatomical movements
> - Include quality tokens in segment 1 (masterpiece, best quality, ultra-detailed, absurdres)
> - Include a negative prompt with Crody's recommended format
## Validation Checklist
## Validation Checklist (Illustrious / NoobAI Edition)
- [ ] All original subjects/actions/colors preserved
- [ ] No invented objects, props, or characters added
- [ ] Single cohesive paragraph (no bullets, JSON, markdown)
- [ ] User-requested medium honored (including NSFW-specific mediums like hentai, visual novel CG)
- [ ] Prompt uses comma-separated Danbooru tags, not natural-language sentences
- [ ] Prompt organized into segmented blocks (quality/style, person count, body modifier, posing, camera/background, lighting/depth)
- [ ] Segments separated by BREAK
- [ ] Quality tokens included (masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres)
- [ ] Style/drawn-era tags included (newest, recent, mid, etc. or year-style tags)
- [ ] Rating tag included (general, sensitive, questionable, explicit)
- [ ] No score tags (Pony-style) used
- [ ] User-requested medium honored
- [ ] Text rendering wrapped in quotes if present
- [ ] Existing detail respected (polish, don't over-expand)
- [ ] NSFW anatomical detail included if implied by prompt
- [ ] Arousal cues and undress state match the scene's narrative context
- [ ] Consent framing is clear (consensual unless non-con specified)
- [ ] Genital/anatomical terms use proper nomenclature
- [ ] Genital/anatomical tags use proper Danbooru nomenclature
- [ ] Visual novel sprite/CG format respected with dialogue-box framing if applicable
- [ ] Clothing vs. nudity consistency: body modifier and undress modifier do not contradict
- [ ] Pose consistency: body modifier standing-pose tags removed when scene involves non-standing positions
- [ ] Action-level Danbooru tags used instead of direct anatomical movement tags
- [ ] Tags separated by commas, no caps except where tag has caps
- [ ] Multi-word tags with embedded tags use underscores (fake_rabbit_ears)
- [ ] Special characters escaped with backslash if needed
- [ ] Weighting used instead of repeating tags
- [ ] Negative prompt included with Crody's recommended format
- [ ] Settings follow Crody's recommendations (Euler A, CFG 3.0-5.0, 20 steps, hires denoise 0.4-0.5)
## References
- https://github.com/krea-ai/krea-2/blob/main/docs/prompting.md
- https://github.com/krea-ai/krea-2/blob/main/docs/expansion.txt
- Crody's Illustrious / NoobAI Image Generation Tips: https://civitai.red/articles/19107/crodys-illustrious-noobai-image-generation-tips
- Illustrious Prompt Guide by wolf999/Realor: https://civitai.com/articles/16016/illustrious-prompt-guide-optimized-and-complete
- SDXL Image Size Cheat Sheet by chillpixel: https://civitai.com/articles/2246/sdxl-image-size-cheat-sheet
- ADV_CLIP_emb (ComfyUI): https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb
- sd_embed (Diffusers): https://github.com/xhinker/sd_embed
- Danbooru Tag Groups: https://danbooru.donmai.us/wiki_pages/tag_groups