LoRa intro.

This is an NSFW-oriented style LoRA.

Many thanks to WAI for the excellent base model 🌷

Please use the base model : WAI-NSFW-illustrious-SDXL (currently trained on v12)

https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl

Target Effects

• Increase maturity of female characters

• Implicitly enhance pornographic elements and intensity by default

• Optimize leg wear features (pantyhose & high heels)

• Preference for "light-colored, natural-toned shaping pantyhose" in hosiery category

• Increase probability of exposed intimate areas (no undergarments)

• Maximize clothing transparency effects when possible

• Prioritize half-body or full-body compositions when no pose tag specified, while increasing co-occurrence probability with male characters

• (Experimental) Explore "alien creature" concepts to improve AI understanding: e.g., crab monster, banana monster, cat-human hybrid, etc.

Tips for NSFWish LoRa v2

• To avoid "too colorful clothes" (e.g. blue, cyan, pink, ...)

  • TRY :

    • neg : [colorful bodywear]

• To increase possibility of natural nude pantyhose

  • TRY :

    • pos : (sheer golden pantyhose)

    • & neg : [[yellow pantyhose]] ( or [yellow pantyhose] ? try it )

• More examples comparing different strength of using the LoRa can see from some of the images I post below

• After testing, working with other LoRa is fine

Hmmm... But, seems it's still got some issues. A new version would be made if I got some spare time...

Suggestions are welcomed !! :)

Recommended Parameters

• Similar to WAI illustrious v12 recommendations, but requires slightly more steps

## Params

Steps: 26-32 (28~30 may be optimal)
CFG scale: 5-7
Sampler: Euler a
Use original dimensions larger than 1024x1024

## Hires

Hires upscale: 1.5
Hires steps: 20
Hires upscaler: R-ESRGAN 4x+ Anime6B
Denoising strength: 0.35 ~ 0.5

Prompts

Positive Prompt Recommendations

masterpiece,best quality,

Negative Prompt Recommendations

bad quality,worst quality,worst detail,sketch,censor,multiple views,

Complex Prompts

• Apply higher weights to:

  • <lora:nsfwish_001:1.2>,(nsfw),(explicit),

  • Consider combining with other NSFW tags

Simplified Prompts

• Use standard weights or experiment with variations:

  • <lora:nsfwish_001:1>,

  • May combine with other NSFW tags

Monster Control Prompts

• Feel free to experiment with similar tags, apply appropriate weights, or consider adding unwanted elements to the negative prompts.

1girl, (human female),
${monster_type} male monster, (or: male ${怪物类别} monster,)

Known Issues

• Initial version is experimental; occasional image distortions observed though overall performance is acceptable

• (Subjective observation) Apparent increased probability of cyan-colored clothing (needs investigation)

• (Subjective observation) "amazing quality" tag may increase distortion risk with simple prompts

LoRa 介绍

这是一个 nsfw 向 style 的 LoRa

非常感谢 WAI 带来的精彩的底模 (base model) 🌷

底模请使用 : WAI-NSFW-illustrious-SDXL (LoRa 在 v12 上训练)

https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl

目标效果

• 增加女孩成熟度

• 貌似默认提升色情元素和程度

• 优化腿部 (丝袜 & 高跟)

• 丝袜类别的偏好是 "浅色系偏自然色调的舍宾袜" (shaping pantyhose)

• 会提高性感部位的出现概率 (无内衣裤)

• 衣物尽可能有透视效果

• 无姿势要求时, 偏向于半身或全身, 且加大出现 male 同图的概率

• (实验) 尝试 "异型生物" 的概念, 让 AI 比较理解这类 "概念" : 比如螃蟹怪, 香蕉怪, 猫怪, ... 等等

参数推荐

• 和 WAI illustrious v12 的推荐类似, 但部署需要稍高些

## 参数

Steps: 26-32 (28~30 may be optimal)
CFG scale: 5-7
Sampler: Euler a
Use original dimensions larger than 1024x1024

## Hires 参数

Hires upscale: 1.5
Hires steps: 20
Hires upscaler: R-ESRGAN 4x+ Anime6B
Denoising strength: 0.35 ~ 0.5

提示词

正向提示词推荐

masterpiece, best quality,

反向提示词推荐

bad quality,worst quality,worst detail,sketch,censor,multiple views,

提示词较多

• 可以给多一些权重 :

  • <lora:nsfwish_001:1.2>,(nsfw),(explicit),

  • 可能可以试试搭配其他 nsfw 的 tag

提示词较简单

• 正常给权重, 或者你可以试试其他权重

  • <lora:nsfwish_001:1>,

  • 可能可以试试搭配其他 nsfw 的 tag

怪物控制提示词推荐

• 类似的 tag 多试试, 可适当给权重, 或者考虑将不要的元素加到 neg prompt

1girl, (human female),
${怪物类别} male monster, (或者: male ${怪物类别} monster,) 

已知问题

• 第一版是实验版本, 感觉还不是太稳定有时会崩图, 但整体效果还行, 可以试玩看看

• (主观感觉) 好像衣服颜色会提升 "青色 (cyan)" 的概率 (需要看看咋回事)

• (主观感觉) amazing quality 似乎会提升简单提示词时的崩图概率