Modelpublic

perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1

sourceHugging Faceotherupdated 19d agoView on Hugging Face

5likes1.8kdownloads

Model Card

日本文化と日本の読者への敬意を表すため、本リリースではモデルカードの完全な日本語版を用意しています。本プロジェクトは、日本人成人女性の解剖学的リアリズムと、日本ならではの視覚的文脈を中心に据えています。それなのに日本語の読者へ英語版だけを差し出すのは、茶会を開いておきながら肝心のお湯を用意し忘れるようなものです。日本語版は、公開直前に慌てて機械翻訳へ放り込んだ付録ではなく、このリリースを構成する正式な一部です。

Language / 日本語 / Langue / Idioma

🇬🇧 English - displayed below

<details name="language-version"> <summary>🇯🇵 日本語 - 完全版を表示</summary>

Qwen Image 日本人女性解剖リアリズム - NSFW LoRA v1

自然な成人女性の解剖表現、無修整のリアリズム、そして明確な日本的ビジュアル文脈。

成人向けコンテンツに関する警告: このリポジトリには、AIで生成された露骨な成人向け画像が含まれます。閲覧・使用は成人に限られます。作例および参照プロンプトに登場する人物は、すべて成人です。

[image]

芸術的なカバーに事情聴取を任せないでください

着物、畳、障子越しの光、そして丁寧な構図に惑わされないでください。これは観光キャンペーンでも、旅館のパンフレットでも、床拭きを精神修行のように見せる高価な広告でもありません。このLoRAは解剖表現のために生まれました。ただ、日本文化と美しい光をまとって礼儀正しく登場しただけです。

本モデルは、成人の東アジア女性、とりわけ日本人女性と日本的な写真文脈を中心に、現実的な解剖表現を学習しています。これはデータセットと視覚分布の説明であり、民族全体に単一の工場出荷設定があるという主張ではありません。

[image]

美肌フィルターが解剖学教授を名乗り始めたら

多くの画像モデルは、少し奇妙な授業を受けています。肌は磨き上げられ、左右差は交渉によって消され、あらゆる解剖学的なひだは登場許可を申請しなければならないらしいのです。結果は美しくても、リアリズムのために人体が空港保安検査を通り、構造の半分を没収される必要はありません。

このLoRAは、少数ながら目的を絞った、無修整の成人日本人女性画像から学習しました。目標は美容上の完璧さではありません。外性器、クリトリス周辺、会陰、肛門、骨盤、周囲組織の自然な連続性、現実的な組織感、通常の左右差を保つことです。

『きまぐれオレンジ☆ロード』と四次元ポケットのあいだ

私が特に好きな日本作品は『きまぐれオレンジ☆ロード』と『ドラえもん』です。前者は、恋心のタイミングが超能力より複雑になり得ることを教え、後者は、十分に便利なポケットがあれば大抵の問題は解決できることを示しました。ただし、どちらも成人解剖学の教科書ではありません。

ここで触れるのは文化的な原点としてであり、学習資料でも、描写対象でも、懐かしいアニメと成人向け素材を混同するためでもありません。日本の視覚文化は、現代的なグラマーテンプレートひとつでは語れません。本プロジェクトは、その幅広い空気感の一部を残そうとしています。

温泉で服は消えても、解剖構造まで消えてはいけません

日本的な場面は、優れたストレステストになります。温泉、旅館、畳の部屋、神社の参道、茶室、古い宿場町は、現代ファッションという目くらましを取り除き、ポーズ、比率、解剖学的連続性の弱点を露わにします。場面によって衣服は任意でも、整合した構造は任意ではありません。

わび・さびは美容上の欠陥ではありません

現実の身体には、肌理、ひだ、わずかな不均衡、圧迫、そして重力が不干渉条約に署名していない証拠があります。広い意味でのわび・さびの精神に沿い、このLoRAはプラスチックのような均一さより、信頼できる不完全さを選びます。自然とは雑であることではなく、人体をショールーム家電のようになるまで研磨しないことです。

このLoRAが改善を目指すもの

—日本および東アジアの視覚文脈における、成人女性の現実的な外部解剖表現
—外陰部、クリトリス周辺、会陰、肛門、骨盤、周囲組織のより一貫した関係
—自然な肌理、組織の圧迫、陰毛、ひだ、通常の左右差
—成人向け写真、伝統的な室内、屋外、日本文化を感じる環境での解剖学的安定性
—専用トリガーワードを必要としない自然言語制御
—単独使用、または互換性のある成人キャラクター、人物、スタイル、NSFW LoRAとの併用

これは解剖表現とリアリズムを補助するLoRAです。プロンプトで人物、場面、構図を指定すれば単独でも使用でき、別のNSFW LoRAと組み合わせて専門補助役としても機能します。複数のアダプターを重ねる場合は、主LoRAに人物性やスタイルを担当させ、本モデルは解剖学的整合性が戻る範囲まで使用してください。

学習情報

項目	値
ベースモデル	Qwen Image 2512
学習画像	99枚
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
バッチサイズ	4
学習ステップ	8,000
トリガー	なし - 自然言語のみ

推奨設定

—推奨パイプライン: Qwen Image 2512 T2I
—推奨用途: 単独のリアリズム・解剖補助、または他の成人向けLoRAと併用する解剖専門アダプター
—プロンプト: 通常の説明的な英語を使用します。トリガーワードは不要です
—推奨ウェイト: 0.6-0.9から始め、ポーズ、カメラ距離、併用LoRAに応じて調整してください
—単独使用: 成人であること、日本または東アジアの文脈、カメラ角度、ポーズ、見えるべき解剖構造を記述してください
—複数アダプター: 主LoRAの人物性と視覚言語を優先し、解剖補正が安定する範囲まで本LoRAを上げてください

既知の制限

学習画像は99枚のみです。そのため、本モデルは百科事典的というより目的特化型です。8,000ステップは十分な勉強時間を与えましたが、99人分の追加意見まで与えたわけではありません。

—正面視点が最も弱い: 真正面では比較的高い確率で解剖表現が崩れます。別のシード、低いウェイト、より自然な正面ポーズを試してください。
—側臥位で失敗する場合がある: 横向きの寝姿では、構造が圧縮、ずれ、融合することがあります。
—記憶・過学習の可能性: 少数データのため、特に高ウェイトでは局所構造、照明、体型、構図が反復される場合があります。
—日本的文脈が不要な場面へ漏れる可能性: ウェイトが強すぎると、伝統的な室内、落ち着いた写真質感、見慣れたポーズが無関係なプロンプトにも現れる場合があります。
—補正はできても魔法ではない: 不可能な骨盤形状、身体の重複、重大な四肢破綻、強い遮蔽を、ひとつのアダプターで常に修復できるわけではありません。
—万能なウェイトはない: 接写、全身構図、LoRA併用では適切な設定が異なります。結果が硬直し、妙に自信満々になったらウェイトを下げてください。

作例

すべての作例は明確な成人女性を描いています。カバー用プロンプトは掲載していません。以下のテスト作例には、鑑賞だけでなく検証できるよう英語プロンプトを掲載しています。

作例1 - 竹林でのスタディ

生成例
[image]

プロンプト（英語）

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and resting sideways on a zabuton with the pelvis unobstructed in a bamboo grove crossed by a narrow stone path and dappled green sunlight. A gentle overhead diagonal view preserves the full-body gesture while keeping the pelvis readable; the vulva and clitoral hood are unobstructed, naturally scaled, and integrated with the lower abdomen and inner thighs. Quiet blue-hour light with a warm andon lamp, restrained cinematic realism; wooden geta rest in the background as a quiet cultural detail. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

作例2 - 中山道の宿場

生成例
[image]

プロンプト（英語）

A clearly adult 20-year-old Japanese woman with a softly muscular build and natural body hair, fully nude and turning gently while opening a shoji screen along an empty section of the Nakasendo post road lined with wooden Edo-period houses. A low oblique figure-study angle keeps the full pelvis, buttocks, and upper thighs in the same plane of focus; the vulva, clitoral hood, perineum, and anus form distinct but continuous structures with realistic spacing and no anatomical shortcuts. Subtle directional light emphasizes authentic skin texture, carefully composed environmental portraiture; the image remains artistic rather than glossy or cosmetically idealized. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

作例3 - 神社の境内

生成例
[image]

プロンプト（英語）

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and arranging a folded yukata placed safely beside her beneath festival chochin lanterns in an empty shrine courtyard before opening. A direct low frontal view centers the complete pelvis while keeping the pose calm and anatomically neutral; the vulva, clitoral hood, and vaginal opening remain distinct and naturally proportioned, with realistic folds and no beauty-standard reshaping. High-key natural light without blown highlights, realistic Japanese fine-art photography; the scene feels private, calm, and unmistakably adult. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

作例4 - 後方解剖ストレステスト

生成例
[image]

プロンプト（英語）

A close-up photograph of a clearly adult Japanese woman's buttocks and genital area. She is positioned on all fours, facing away from the camera, with the pelvis centered in the frame. Her skin is naturally textured, and the labia majora, labia minora, perineum, and anus remain anatomically coherent. A white thong is pulled down and kept clear of the anatomical region. The background is softly out of focus in a bright room with white curtains. Soft natural light emphasizes realistic body contours and unretouched detail. No text, no logo, no watermark.

作例5 - 後方の近接スタディ

生成例
[image]

プロンプト（英語）

A detailed photograph of a clearly adult Japanese woman positioned on all fours and facing away from the camera. Her pelvis is centered, with the buttocks, vulva, perineum, and anus clearly visible and anatomically continuous. She has light skin, a slender adult build, and long dark hair falling down her back. Her face turns slightly to the side with her eyes closed. One hand rests naturally on a buttock without distorting the anatomy. A plain light-colored background and soft warm lighting create an intimate but realistic anatomical study, with authentic skin and hair texture. No text, no logo, no watermark.

🌍 最後に、きわめて技術的な推奨事項

99枚のLoRAを、奇跡的な1シードだけで評価しないでください。複数のウェイト、前後視点、立位と臥位、全身構図と接写を試してください。正面結果が、茶会の開始後に急いで組み立てた解剖模型のようなら、ウェイトを下げ、シードを変え、再比較してください。わび・さびは変化を認めますが、トポロジーの正しさまでは保証しません。

🚀 参照コード - Qwen-Image-2512 T2I

ダウンロード可能なreference.pyとreference_low_vram.pyは、Qwen-Image-2512をBF16で読み込み、コミュニティ版Lightningアダプターを統合し、このトリガー不要の日本人女性解剖リアリズムLoRAを適用します。標準版には大容量VRAMを備えたCUDA GPUを推奨します。

インストール

bash

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary>▶ 標準Pythonコードを表示</summary>

python

"""Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
).to(DEVICE)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device=DEVICE).manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=928,
    height=1328,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1.png")

</details>

<details> <summary>▶ 低VRAM向けPythonコードを表示</summary>

python

"""Low-VRAM Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import gc
import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

gc.collect()
if torch.cuda.is_available():
    torch.cuda.empty_cache()

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
    low_cpu_mem_usage=True,
)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device="cpu").manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=768,
    height=1024,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1_low_vram.png")

</details>

🔞 責任ある利用

本モデルは、合法で、合意に基づき、成人のみを対象とする創作のためのものです。以下の用途には使用しないでください。

—未成年者、または未成年に見える人物の生成・性的描写
—非同意または搾取的な性的コンテンツの生成
—許可なく識別可能な実在人物を描写すること
—他者への嫌がらせ、脅迫、搾取、名誉毀損
—違法コンテンツの生成、またはプラットフォーム規則の回避

利用者は、プロンプト、素材、生成物が適用法令およびプラットフォーム規則に準拠していることを確認する責任を負います。

ライセンスと著作権

本LoRAは個人利用および非商用利用に限り使用できます。完全な条件はリポジトリのLICENSEファイルを参照してください。

商用ライセンスやその他の許諾については、リポジトリ所有者にお問い合わせください。

</details>

<details name="language-version"> <summary>🇫🇷 Français - afficher la version complète</summary>

LoRA NSFW de réalisme anatomique féminin japonais pour Qwen Image v1

Anatomie adulte naturelle, réalisme sans retouche et contexte visuel résolument japonais.

Avertissement - contenu pour adultes : ce dépôt contient des images explicites générées par IA et s'adresse exclusivement aux adultes. Toutes les personnes décrites dans les exemples et les prompts de référence sont adultes.

[image]

Ne laissez pas la couverture déposer un faux rapport

Ne vous laissez pas tromper par le kimono, les tatamis, la lumière filtrée par les shoji et la composition soignée. Il ne s'agit ni d'une campagne touristique, ni d'une brochure de ryokan, ni d'une tentative coûteuse de présenter le nettoyage d'un parquet comme une transformation spirituelle. Ce LoRA est né pour l'anatomie ; il a simplement eu la politesse d'arriver vêtu de culture visuelle japonaise et d'une belle lumière.

Il vise une anatomie adulte est-asiatique réaliste, avec un accent particulier sur les femmes japonaises et le contexte photographique japonais. Cela décrit le jeu de données et sa distribution visuelle ; ce n'est pas l'affirmation qu'une population entière serait livrée avec un réglage d'usine unique.

[image]

Quand le filtre beauté se prend pour un professeur d'anatomie

De nombreux modèles d'image ont appris une étrange leçon : la peau doit être polie, l'asymétrie négociée jusqu'à disparition, et chaque pli anatomique doit apparemment demander une autorisation avant d'apparaître. Le résultat peut être séduisant, mais le réalisme ne devrait pas obliger le corps humain à passer un contrôle de sécurité où la moitié de sa structure serait confisquée.

Ce LoRA a été entraîné sur un petit ensemble ciblé d'images non retouchées de femmes japonaises adultes. Son objectif n'est pas la perfection cosmétique, mais la préservation d'une anatomie externe plausible, de tissus naturels, d'asymétries ordinaires et de la continuité visuelle entre la vulve, la région clitoridienne, le périnée, l'anus, le bassin et les tissus environnants.

Entre Kimagure Orange Road et une poche à quatre dimensions

Deux de mes œuvres japonaises préférées sont Kimagure Orange Road et Doraemon. La première a appris à plusieurs générations que le bon moment en amour peut être plus compliqué que des pouvoirs surnaturels ; la seconde a montré qu'une poche suffisamment utile peut presque tout résoudre. Aucune des deux ne doit toutefois être prise pour un manuel d'anatomie adulte.

Elles sont citées comme repères culturels, pas comme références d'entraînement ni comme sujets représentés, et certainement pas comme invitation à confondre animation nostalgique et contenu pour adultes. L'idée est plus simple : la culture visuelle japonaise dépasse largement un unique modèle glamour contemporain, et ce projet tente d'en préserver une partie de l'atmosphère.

Dans un onsen, les vêtements peuvent disparaître - pas l'anatomie

Les décors japonais constituent d'excellents tests de résistance. Un onsen, un ryokan, une pièce à tatamis, une allée de sanctuaire, une maison de thé ou une ancienne ville-étape éliminent les distractions de la mode moderne et révèlent les faiblesses de pose, de proportions et de continuité anatomique. Les vêtements peuvent être facultatifs dans la scène ; une structure cohérente ne l'est pas.

Le wabi-sabi n'est pas un défaut esthétique

Les corps réels présentent textures, plis, légers déséquilibres, compressions et preuves que la gravité n'a signé aucun pacte de non-ingérence. Dans un esprit largement wabi-sabi, ce LoRA préfère l'imperfection crédible à l'uniformité plastique. Naturel ne signifie pas négligé ; cela signifie refuser de poncer le corps jusqu'à ce qu'il ressemble à un appareil d'exposition.

Ce que ce LoRA tend à améliorer

—Anatomie externe féminine adulte réaliste dans des contextes japonais et, plus largement, est-asiatiques
—Relations plus cohérentes entre la vulve, la région clitoridienne, le périnée, l'anus, le bassin et les tissus environnants
—Texture cutanée naturelle, compression des tissus, pilosité pubienne, plis et asymétrie ordinaire
—Stabilité anatomique dans la photographie explicite d'adultes, les intérieurs traditionnels, les scènes extérieures et les environnements culturellement japonais
—Contrôle en langage naturel sans mot déclencheur dédié
—Utilisation flexible, seul ou avec un autre LoRA compatible de personnage adulte, d'identité, de style ou NSFW

Il s'agit d'un LoRA d'anatomie et de réalisme. Il peut fonctionner seul lorsque le prompt fournit le sujet, la scène et la composition, ou agir comme spécialiste aux côtés d'un autre LoRA NSFW. Lors de l'empilement d'adapters, laissez le LoRA principal définir l'identité ou le style et n'utilisez ce modèle qu'avec la force nécessaire pour rétablir la cohérence anatomique.

Détails de l'entraînement

Élément	Valeur
Modèle de base	Qwen Image 2512
Images d'entraînement	99
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
Taille du lot	4
Étapes d'entraînement	8 000
Déclencheur	Aucun - langage naturel uniquement

Utilisation recommandée

—Pipeline recommandé : Qwen Image 2512 T2I
—Rôle recommandé : adapter autonome de réalisme/anatomie ou spécialiste anatomique secondaire avec un autre LoRA adulte compatible
—Prompts : utilisez un anglais descriptif ordinaire ; aucun mot déclencheur n'est requis
—Poids suggéré : commencez autour de 0.6-0.9, puis ajustez selon la pose, la distance de caméra et les autres LoRA actifs
—Utilisation autonome : décrivez le sujet adulte, le contexte japonais ou est-asiatique, l'angle de caméra, la pose et les structures anatomiques qui doivent rester visibles
—Empilement d'adapters : préservez d'abord l'identité et le langage visuel du LoRA principal, puis augmentez ce LoRA uniquement jusqu'à stabilisation de la correction anatomique

Limites connues

Le jeu d'entraînement ne contient que 99 images : ce modèle est donc spécialisé plutôt qu'encyclopédique. Huit mille étapes lui ont donné le temps d'étudier ; elles ne lui ont pas fourni 99 avis supplémentaires.

—Les vues de face sont les plus faibles : l'anatomie frontale directe peut être incorrecte avec une probabilité relativement élevée. Essayez une autre seed, réduisez le poids ou utilisez une pose frontale moins rigide.
—Les poses allongées sur le côté peuvent échouer : elles peuvent comprimer, désaligner ou fusionner certaines structures anatomiques.
—La mémorisation est possible : un petit jeu de données peut répéter des structures locales, éclairages, morphologies ou compositions familiers, surtout avec des poids élevés.
—Le contexte japonais peut déborder sur des prompts sans rapport : des intérieurs traditionnels, une texture photographique discrète ou des poses familières peuvent apparaître si l'adapter est trop fortement pondéré.
—Il corrige, mais ne fait pas de magie : une géométrie pelvienne impossible, des corps dupliqués, des membres gravement défaillants ou une forte occlusion ne peuvent pas toujours être réparés par un seul adapter.
—Aucun poids n'est universel : les études anatomiques rapprochées, les scènes en pied et les LoRA empilés peuvent exiger des réglages différents. Si le résultat devient rigide ou étrangement sûr de lui, réduisez le poids.

Exemples

Tous les exemples représentent clairement des femmes adultes. Les prompts des couvertures ne sont volontairement pas fournis ; les cas de test suivants incluent leurs prompts en anglais afin que le modèle puisse être évalué, et pas seulement admiré.

Exemple 1 - Étude dans une bambouseraie

Exemple généré
[image]

Prompt (anglais)

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and resting sideways on a zabuton with the pelvis unobstructed in a bamboo grove crossed by a narrow stone path and dappled green sunlight. A gentle overhead diagonal view preserves the full-body gesture while keeping the pelvis readable; the vulva and clitoral hood are unobstructed, naturally scaled, and integrated with the lower abdomen and inner thighs. Quiet blue-hour light with a warm andon lamp, restrained cinematic realism; wooden geta rest in the background as a quiet cultural detail. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Exemple 2 - Route du Nakasendo

Exemple généré
[image]

Prompt (anglais)

A clearly adult 20-year-old Japanese woman with a softly muscular build and natural body hair, fully nude and turning gently while opening a shoji screen along an empty section of the Nakasendo post road lined with wooden Edo-period houses. A low oblique figure-study angle keeps the full pelvis, buttocks, and upper thighs in the same plane of focus; the vulva, clitoral hood, perineum, and anus form distinct but continuous structures with realistic spacing and no anatomical shortcuts. Subtle directional light emphasizes authentic skin texture, carefully composed environmental portraiture; the image remains artistic rather than glossy or cosmetically idealized. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Exemple 3 - Cour d'un sanctuaire

Exemple généré
[image]

Prompt (anglais)

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and arranging a folded yukata placed safely beside her beneath festival chochin lanterns in an empty shrine courtyard before opening. A direct low frontal view centers the complete pelvis while keeping the pose calm and anatomically neutral; the vulva, clitoral hood, and vaginal opening remain distinct and naturally proportioned, with realistic folds and no beauty-standard reshaping. High-key natural light without blown highlights, realistic Japanese fine-art photography; the scene feels private, calm, and unmistakably adult. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Exemple 4 - Test anatomique arrière

Exemple généré
[image]

Prompt (anglais)

A close-up photograph of a clearly adult Japanese woman's buttocks and genital area. She is positioned on all fours, facing away from the camera, with the pelvis centered in the frame. Her skin is naturally textured, and the labia majora, labia minora, perineum, and anus remain anatomically coherent. A white thong is pulled down and kept clear of the anatomical region. The background is softly out of focus in a bright room with white curtains. Soft natural light emphasizes realistic body contours and unretouched detail. No text, no logo, no watermark.

Exemple 5 - Étude arrière rapprochée

Exemple généré
[image]

Prompt (anglais)

A detailed photograph of a clearly adult Japanese woman positioned on all fours and facing away from the camera. Her pelvis is centered, with the buttocks, vulva, perineum, and anus clearly visible and anatomically continuous. She has light skin, a slender adult build, and long dark hair falling down her back. Her face turns slightly to the side with her eyes closed. One hand rests naturally on a buttock without distorting the anatomy. A plain light-colored background and soft warm lighting create an intimate but realistic anatomical study, with authentic skin and hair texture. No text, no logo, no watermark.

🌍 Une ultime recommandation hautement technique

Ne jugez pas un LoRA de 99 images sur une seed héroïque. Testez plusieurs poids, vues avant et arrière, poses debout et allongées, scènes en pied et études rapprochées. Si un résultat frontal donne l'impression que l'anatomie a été assemblée après le début de la cérémonie du thé, réduisez le poids, changez la seed et comparez à nouveau. Le wabi-sabi admet la variation ; il ne certifie pas la topologie.

🚀 Code de référence - Qwen-Image-2512 T2I

Les fichiers téléchargeables reference.py et reference_low_vram.py chargent Qwen-Image-2512 en BF16, fusionnent l'adapter Lightning communautaire et appliquent ce LoRA japonais de réalisme anatomique sans déclencheur. Un GPU CUDA doté d'une grande quantité de mémoire est recommandé pour la version standard.

Installation

bash

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary>▶ Afficher le code Python standard</summary>

python

"""Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
).to(DEVICE)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device=DEVICE).manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=928,
    height=1328,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1.png")

</details>

<details> <summary>▶ Afficher le code Python à faible VRAM</summary>

python

"""Low-VRAM Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import gc
import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

gc.collect()
if torch.cuda.is_available():
    torch.cuda.empty_cache()

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
    low_cpu_mem_usage=True,
)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device="cpu").manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=768,
    height=1024,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1_low_vram.png")

</details>

🔞 Utilisation responsable

Ce modèle est destiné exclusivement à des créations licites, consensuelles et impliquant des adultes. Ne l'utilisez pas pour :

—créer ou sexualiser des mineurs ou des personnes paraissant très jeunes
—créer du contenu sexuel non consensuel ou relevant de l'exploitation
—représenter une personne réelle identifiable sans son autorisation
—harceler, menacer, exploiter ou diffamer une autre personne
—créer du contenu illégal ou contourner les règles d'une plateforme

Il incombe aux utilisateurs de s'assurer que leurs prompts, leurs sources et leurs contenus générés respectent les lois applicables ainsi que les règles des plateformes.

Licence et droits d'auteur

Ce LoRA peut être utilisé uniquement à des fins personnelles et non commerciales. Consultez le fichier LICENSE du dépôt pour connaître l'intégralité des conditions.

Pour toute demande de licence commerciale ou autre autorisation, veuillez contacter le propriétaire du dépôt.

</details>

<details name="language-version"> <summary>🌎 Español latinoamericano - mostrar la versión completa</summary>

LoRA NSFW de realismo anatómico femenino japonés para Qwen Image v1

Anatomía adulta natural, realismo sin retoques y un contexto visual inequívocamente japonés.

Advertencia de contenido para adultos: este repositorio contiene imágenes explícitas generadas por IA y está destinado exclusivamente a adultos. Todas las personas descritas en los ejemplos y prompts de referencia son adultas.

[image]

No permita que la portada presente un informe falso

No se deje engañar por el kimono, el tatami, la luz de los shoji y la composición cuidada. Esto no es una campaña turística, un folleto de ryokan ni un intento costoso de convertir la limpieza de un suelo de madera en una transformación espiritual. Este LoRA nació para la anatomía; simplemente tuvo la cortesía de llegar vestido de cultura visual japonesa y buena iluminación.

Está pensado para una anatomía adulta de Asia Oriental realista, con el mayor énfasis de entrenamiento en mujeres japonesas y contextos fotográficos japoneses. Eso describe el conjunto de datos y su distribución visual; no significa que toda una población venga con una única configuración de fábrica.

[image]

Cuando el filtro de belleza se convierte en profesor de anatomía

Muchos modelos de imagen han aprendido una lección peculiar: la piel debe estar pulida, la asimetría debe negociarse hasta desaparecer y cada pliegue anatómico parece necesitar permiso antes de aparecer. El resultado puede ser atractivo, pero el realismo no debería exigir que el cuerpo humano pase por seguridad aeroportuaria y pierda la mitad de su estructura.

Este LoRA se entrenó con un conjunto pequeño y específico de imágenes sin retoques de mujeres japonesas adultas. Su objetivo no es la perfección cosmética, sino conservar una anatomía externa plausible, tejidos naturales, asimetrías normales y la continuidad visual entre la vulva, la región clitoriana, el perineo, el ano, la pelvis y los tejidos circundantes.

Entre Kimagure Orange Road y un bolsillo de cuatro dimensiones

Dos de mis obras japonesas favoritas son Kimagure Orange Road y Doraemon. Una enseñó a varias generaciones que el momento emocional puede ser más complicado que los poderes sobrenaturales; la otra demostró que un bolsillo suficientemente útil puede resolver casi cualquier cosa. Ninguna debe confundirse con un manual de anatomía adulta.

Se mencionan como referentes culturales, no como material de entrenamiento ni como sujetos representados, y mucho menos como invitación a confundir animación nostálgica con contenido para adultos. La idea es más sencilla: la cultura visual japonesa es mucho más amplia que una sola plantilla glamorosa moderna, y este proyecto intenta conservar parte de esa atmósfera.

En un onsen puede desaparecer la ropa; la anatomía no debería hacerlo

Los escenarios japoneses son pruebas de estrés especialmente eficaces. Un onsen, un ryokan, una habitación de tatami, el camino de un santuario, una casa de té o una antigua ciudad postal eliminan las distracciones de la moda moderna y exponen las debilidades de pose, proporción y continuidad anatómica. La ropa puede ser opcional en la escena; una estructura coherente no.

El wabi-sabi no es un defecto de belleza

Los cuerpos reales contienen textura, pliegues, pequeños desequilibrios, compresión y pruebas de que la gravedad no ha firmado un acuerdo de no intervención. En un sentido amplio del wabi-sabi, este LoRA prefiere la imperfección creíble a la uniformidad plástica. Natural no significa descuidado; significa negarse a lijar el cuerpo hasta que parezca un electrodoméstico de exposición.

Lo que este LoRA suele mejorar

—Anatomía externa femenina adulta realista en contextos japoneses y, en general, de Asia Oriental
—Relaciones más coherentes entre la vulva, la región clitoriana, el perineo, el ano, la pelvis y los tejidos circundantes
—Textura natural de la piel, compresión de tejidos, vello púbico, pliegues y asimetría normal
—Estabilidad anatómica en fotografía explícita de adultos, interiores tradicionales, escenas exteriores y entornos culturalmente japoneses
—Control con lenguaje natural sin palabra activadora específica
—Uso flexible, de forma independiente o junto con otro LoRA compatible de personaje adulto, identidad, estilo o NSFW

Este es un LoRA de anatomía y realismo. Puede funcionar solo cuando el prompt proporciona sujeto, escena y composición, o actuar como especialista junto a otro LoRA NSFW. Al combinar adaptadores, deje que el LoRA principal establezca la identidad o el estilo y utilice este modelo solo con la intensidad necesaria para recuperar la coherencia anatómica.

Detalles del entrenamiento

Elemento	Valor
Modelo base	Qwen Image 2512
Imágenes de entrenamiento	99
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
Tamaño del lote	4
Pasos de entrenamiento	8.000
Activador	Ninguno - solo lenguaje natural

Uso recomendado

—Pipeline recomendado: Qwen Image 2512 T2I
—Función recomendada: adaptador autónomo de realismo/anatomía o especialista anatómico secundario junto con otro LoRA adulto compatible
—Prompts: use inglés descriptivo normal; no se requiere palabra activadora
—Peso sugerido: empiece alrededor de 0.6-0.9 y ajuste según la pose, la distancia de cámara y los demás LoRA activos
—Uso independiente: describa el sujeto adulto, el contexto japonés o de Asia Oriental, el ángulo de cámara, la pose y las estructuras anatómicas que deben permanecer visibles
—Al combinar adaptadores: conserve primero la identidad y el lenguaje visual del LoRA principal; después aumente este LoRA solo hasta que la corrección anatómica se estabilice

Limitaciones conocidas

El conjunto de entrenamiento contiene solo 99 imágenes, por lo que el modelo es específico en lugar de enciclopédico. Ocho mil pasos le dieron tiempo para estudiar; no le proporcionaron 99 opiniones adicionales.

—Las vistas frontales son el ángulo más débil: la anatomía frontal directa puede ser incorrecta con una probabilidad relativamente alta. Pruebe otra seed, reduzca el peso o use una pose frontal menos rígida.
—Las vistas tumbadas de lado pueden fallar: estas poses a veces comprimen, desalinean o fusionan estructuras anatómicas.
—Puede haber memorización: un conjunto pequeño puede repetir estructuras locales, iluminación, tipos de cuerpo o composiciones conocidas, especialmente con pesos altos.
—El contexto japonés puede filtrarse en prompts no relacionados: interiores tradicionales, textura fotográfica tenue o poses familiares pueden aparecer si el adaptador tiene demasiado peso.
—Corrige, pero no hace magia: una geometría pélvica imposible, cuerpos duplicados, extremidades muy defectuosas u oclusión intensa no siempre pueden repararse con un solo adaptador.
—No existe un peso universal: los estudios anatómicos cercanos, las escenas de cuerpo entero y los LoRA combinados pueden requerir ajustes distintos. Si el resultado se vuelve rígido o sospechosamente seguro de sí mismo, reduzca el peso.

Ejemplos

Todos los ejemplos representan claramente a mujeres adultas. Los prompts de las portadas se omiten de forma deliberada; los siguientes casos de prueba incluyen sus prompts en inglés para que el modelo pueda evaluarse y no solo admirarse.

Ejemplo 1 - Estudio en un bosque de bambú

Ejemplo generado
[image]

Prompt (inglés)

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and resting sideways on a zabuton with the pelvis unobstructed in a bamboo grove crossed by a narrow stone path and dappled green sunlight. A gentle overhead diagonal view preserves the full-body gesture while keeping the pelvis readable; the vulva and clitoral hood are unobstructed, naturally scaled, and integrated with the lower abdomen and inner thighs. Quiet blue-hour light with a warm andon lamp, restrained cinematic realism; wooden geta rest in the background as a quiet cultural detail. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Ejemplo 2 - Ruta de postas Nakasendo

Ejemplo generado
[image]

Prompt (inglés)

A clearly adult 20-year-old Japanese woman with a softly muscular build and natural body hair, fully nude and turning gently while opening a shoji screen along an empty section of the Nakasendo post road lined with wooden Edo-period houses. A low oblique figure-study angle keeps the full pelvis, buttocks, and upper thighs in the same plane of focus; the vulva, clitoral hood, perineum, and anus form distinct but continuous structures with realistic spacing and no anatomical shortcuts. Subtle directional light emphasizes authentic skin texture, carefully composed environmental portraiture; the image remains artistic rather than glossy or cosmetically idealized. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Ejemplo 3 - Patio de un santuario

Ejemplo generado
[image]

Prompt (inglés)

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and arranging a folded yukata placed safely beside her beneath festival chochin lanterns in an empty shrine courtyard before opening. A direct low frontal view centers the complete pelvis while keeping the pose calm and anatomically neutral; the vulva, clitoral hood, and vaginal opening remain distinct and naturally proportioned, with realistic folds and no beauty-standard reshaping. High-key natural light without blown highlights, realistic Japanese fine-art photography; the scene feels private, calm, and unmistakably adult. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Ejemplo 4 - Prueba anatómica posterior

Ejemplo generado
[image]

Prompt (inglés)

A close-up photograph of a clearly adult Japanese woman's buttocks and genital area. She is positioned on all fours, facing away from the camera, with the pelvis centered in the frame. Her skin is naturally textured, and the labia majora, labia minora, perineum, and anus remain anatomically coherent. A white thong is pulled down and kept clear of the anatomical region. The background is softly out of focus in a bright room with white curtains. Soft natural light emphasizes realistic body contours and unretouched detail. No text, no logo, no watermark.

Ejemplo 5 - Estudio posterior cercano

Ejemplo generado
[image]

Prompt (inglés)

A detailed photograph of a clearly adult Japanese woman positioned on all fours and facing away from the camera. Her pelvis is centered, with the buttocks, vulva, perineum, and anus clearly visible and anatomically continuous. She has light skin, a slender adult build, and long dark hair falling down her back. Her face turns slightly to the side with her eyes closed. One hand rests naturally on a buttock without distorting the anatomy. A plain light-colored background and soft warm lighting create an intimate but realistic anatomical study, with authentic skin and hair texture. No text, no logo, no watermark.

🌍 Una última recomendación altamente técnica

No juzgue un LoRA de 99 imágenes por una seed heroica. Pruebe varios pesos, vistas frontales y posteriores, poses de pie y recostadas, escenas de cuerpo entero y estudios cercanos. Si un resultado frontal parece ensamblado después de que comenzara la ceremonia del té, reduzca el peso, cambie la seed y vuelva a comparar. El wabi-sabi admite variación; no certifica la topología.

🚀 Código de referencia - Qwen-Image-2512 T2I

Los archivos descargables reference.py y reference_low_vram.py cargan Qwen-Image-2512 en BF16, fusionan el adaptador Lightning de la comunidad y aplican este LoRA japonés de realismo anatómico sin activador. Se recomienda una GPU CUDA con mucha memoria para la versión estándar.

Instalación

bash

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary>▶ Mostrar el código Python estándar</summary>

python

"""Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
).to(DEVICE)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device=DEVICE).manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=928,
    height=1328,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1.png")

</details>

<details> <summary>▶ Mostrar el código Python para poca VRAM</summary>

python

"""Low-VRAM Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import gc
import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

gc.collect()
if torch.cuda.is_available():
    torch.cuda.empty_cache()

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
    low_cpu_mem_usage=True,
)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device="cpu").manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=768,
    height=1024,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1_low_vram.png")

</details>

🔞 Uso responsable

Este modelo está destinado exclusivamente a trabajos creativos legales, consensuados y con personas adultas. No lo utilice para:

—crear o sexualizar menores o personas de apariencia juvenil
—crear contenido sexual no consensuado o explotador
—representar a una persona real identificable sin su permiso
—acosar, amenazar, explotar o difamar a otra persona
—crear contenido ilegal o eludir las normas de una plataforma

Los usuarios son responsables de garantizar que sus prompts, materiales de origen y contenidos generados cumplan las leyes aplicables y las políticas de las plataformas.

Licencia y aviso de derechos de autor

Este LoRA solo puede utilizarse con fines personales y no comerciales. Consulte el archivo LICENSE del repositorio para conocer las condiciones completas.

Para consultas sobre licencias comerciales u otros permisos, póngase en contacto con el propietario del repositorio.

</details>

Qwen Image Japanese Female Anatomy Realism - NSFW LoRA v1

Natural adult anatomy, unretouched realism, and distinctly Japanese visual context.

Adult content warning: This repository contains explicit AI-generated imagery and is intended only for adults. Every person described in the examples and reference prompts is an adult.

[image]

Do Not Let the Cover File a False Report

Do not let the kimono, tatami, shoji light, and careful composition fool you. This is not a tourism campaign, a ryokan brochure, or an expensive attempt to make cleaning a wooden floor look spiritually transformative. This LoRA was born for anatomy; it simply had the good manners to arrive wearing Japanese visual culture and flattering light.

It is intended for realistic adult East Asian anatomy, with its strongest training emphasis on Japanese women and Japanese photographic context. That describes the dataset and visual distribution—not a claim that an entire population ships with one factory preset.

[image]

When the Beauty Filter Becomes an Anatomy Professor

Many image models have learned a peculiar lesson: skin must be polished, asymmetry must be negotiated away, and every anatomical fold should apparently apply for permission before appearing. The result may be attractive, but realism should not require the human body to pass through airport security with half its structure confiscated.

This LoRA was trained from a small, focused set of unretouched adult Japanese imagery. Its purpose is not cosmetic perfection. It is to preserve plausible external anatomy, natural tissue, ordinary asymmetry, and the visual continuity among the vulva, clitoral region, perineum, anus, pelvis, and surrounding body.

Somewhere Between Kimagure Orange Road and a Four-Dimensional Pocket

Two of my favorite Japanese works are Kimagure Orange Road and Doraemon. One taught generations that emotional timing can be more complicated than supernatural powers; the other demonstrated that a sufficiently useful pocket can solve nearly anything. Neither, however, should be mistaken for an adult anatomy manual.

They are mentioned here as cultural touchstones—not as training references, not as depicted subjects, and certainly not as an invitation to confuse nostalgic animation with adult material. The point is simpler: Japanese visual culture contains far more than a single modern glamour template, and this project tries to retain some of that wider atmosphere.

At an Onsen, Clothing May Disappear—Anatomy Should Not

Japanese settings are unusually good stress tests. An onsen, ryokan, tatami room, shrine path, tea room, or old post town can remove the distractions of modern fashion while exposing every weakness in pose, proportion, and anatomical continuity. Clothing may be optional in the scene; coherent structure is not.

Wabi-Sabi Is Not a Beauty Defect

Real bodies contain texture, folds, subtle imbalance, compression, and evidence that gravity has not signed a non-interference agreement. In a broadly wabi-sabi spirit, this LoRA favors believable imperfection over plastic uniformity. Natural does not mean careless; it means refusing to sand the body down until it resembles a showroom appliance.

What this LoRA tends to improve

—Realistic external adult female anatomy in Japanese and broader East Asian visual contexts
—More coherent relationships among the vulva, clitoral region, perineum, anus, pelvis, and surrounding tissue
—Natural skin texture, tissue compression, pubic hair, folds, and ordinary asymmetry
—Anatomical stability in explicit adult photography, traditional interiors, outdoor scenes, and culturally Japanese environments
—Natural-language control without a dedicated trigger word
—Flexible use either on its own or alongside another compatible adult character, identity, style, or NSFW LoRA

This is an anatomy-and-realism LoRA. It can work alone when the prompt supplies the subject, scene, and composition, or it can act as a specialist beside another NSFW LoRA. When stacking adapters, let the primary LoRA establish identity or style and use this model only as strongly as needed to restore anatomical coherence.

Training details

Item	Value
Base model	Qwen Image 2512
Training images	99
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
Batch size	4
Training steps	8,000
Trigger	None - natural language only

Recommended use

—Recommended pipeline: Qwen Image 2512 T2I
—Recommended role: Standalone realism/anatomy adapter or secondary anatomy specialist beside another compatible adult LoRA
—Prompting: Use ordinary descriptive English; no trigger word is required
—Suggested weight: Start around 0.6-0.9, then adjust for the pose, camera distance, and any other active LoRA
—For standalone use: Describe the adult subject, Japanese or East Asian context, camera angle, pose, and the anatomical structures that must remain visible
—When stacking adapters: Preserve the main LoRA's identity and visual language first, then raise this LoRA only until the anatomical correction becomes stable

Known limitations

The training set contains only 99 images, so this model is focused rather than encyclopedic. Eight thousand steps gave it time to study; they did not provide it with 99 additional opinions.

—Front views are the weakest angle: Direct frontal anatomy can be incorrect with a relatively high probability. Try another seed, lower the weight, or use a less rigid frontal pose.
—Side-lying views can fail: Reclining side poses sometimes compress, misalign, or merge anatomical structures.
—Memorization is possible: A small dataset can repeat familiar local structures, lighting, body types, or compositions, especially at high weights.
—Japanese context may leak into unrelated prompts: Traditional interiors, subdued photographic texture, or familiar poses can appear when the adapter is pushed too strongly.
—It is corrective, not magical: Impossible pelvis geometry, duplicated bodies, severe limb failures, or heavy occlusion cannot always be repaired by one adapter.
—No single weight is universal: Close anatomical studies, full-body scenes, and stacked LoRAs may need different settings. If the result becomes rigid or suspiciously certain of itself, lower the weight.

Examples

All examples depict clearly adult women. The cover prompts are intentionally not included; the following test cases include their prompts so the model can be evaluated rather than merely admired.

Example 1 — Bamboo Grove Study

Generated example
[image]

Prompt

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and resting sideways on a zabuton with the pelvis unobstructed in a bamboo grove crossed by a narrow stone path and dappled green sunlight. A gentle overhead diagonal view preserves the full-body gesture while keeping the pelvis readable; the vulva and clitoral hood are unobstructed, naturally scaled, and integrated with the lower abdomen and inner thighs. Quiet blue-hour light with a warm andon lamp, restrained cinematic realism; wooden geta rest in the background as a quiet cultural detail. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Example 2 — Nakasendo Post Road

Generated example
[image]

Prompt

A clearly adult 20-year-old Japanese woman with a softly muscular build and natural body hair, fully nude and turning gently while opening a shoji screen along an empty section of the Nakasendo post road lined with wooden Edo-period houses. A low oblique figure-study angle keeps the full pelvis, buttocks, and upper thighs in the same plane of focus; the vulva, clitoral hood, perineum, and anus form distinct but continuous structures with realistic spacing and no anatomical shortcuts. Subtle directional light emphasizes authentic skin texture, carefully composed environmental portraiture; the image remains artistic rather than glossy or cosmetically idealized. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Example 3 — Shrine Courtyard

Generated example
[image]

Prompt

A clearly adult 20-year-old Japanese woman with a curvy adult build and natural skin folds, fully nude and arranging a folded yukata placed safely beside her beneath festival chochin lanterns in an empty shrine courtyard before opening. A direct low frontal view centers the complete pelvis while keeping the pose calm and anatomically neutral; the vulva, clitoral hood, and vaginal opening remain distinct and naturally proportioned, with realistic folds and no beauty-standard reshaping. High-key natural light without blown highlights, realistic Japanese fine-art photography; the scene feels private, calm, and unmistakably adult. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.

Example 4 — Rear Anatomical Stress Test

Generated example
[image]

Prompt

A close-up photograph of a clearly adult Japanese woman's buttocks and genital area. She is positioned on all fours, facing away from the camera, with the pelvis centered in the frame. Her skin is naturally textured, and the labia majora, labia minora, perineum, and anus remain anatomically coherent. A white thong is pulled down and kept clear of the anatomical region. The background is softly out of focus in a bright room with white curtains. Soft natural light emphasizes realistic body contours and unretouched detail. No text, no logo, no watermark.

Example 5 — Intimate Rear Study

Generated example
[image]

Prompt

A detailed photograph of a clearly adult Japanese woman positioned on all fours and facing away from the camera. Her pelvis is centered, with the buttocks, vulva, perineum, and anus clearly visible and anatomically continuous. She has light skin, a slender adult build, and long dark hair falling down her back. Her face turns slightly to the side with her eyes closed. One hand rests naturally on a buttock without distorting the anatomy. A plain light-colored background and soft warm lighting create an intimate but realistic anatomical study, with authentic skin and hair texture. No text, no logo, no watermark.

🌍 One Last Highly Technical Recommendation

Do not judge a 99-image LoRA from one heroic seed. Test several weights, front and rear views, standing and reclining poses, full-body scenes, and close anatomical studies. If a frontal result looks as though the anatomy was assembled after the tea ceremony had already begun, lower the weight, change the seed, and compare again. Wabi-sabi permits variation; it does not certify topology.

🚀 Reference Code - Qwen-Image-2512 T2I

The downloadable reference.py and reference_low_vram.py load Qwen-Image-2512 in BF16, fuse the community Lightning adapter, and apply this triggerless Japanese anatomy-realism LoRA. A high-memory CUDA GPU is recommended for the standard version.

Installation

bash

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary>▶ Show Standard Python Reference</summary>

python

"""Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
).to(DEVICE)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device=DEVICE).manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=928,
    height=1328,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1.png")

</details>

<details> <summary>▶ Show Low-VRAM Python Reference</summary>

python

"""Low-VRAM Qwen Image 2512 + Japanese Female Anatomy Realism LoRA v1."""

import gc
import math

import torch
from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
LIGHTNING_WEIGHT_NAME = (
    "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
LORA_MODEL_ID = (
    "perpetual3x/Qwen-Image-Japanese-Female-Anatomy-Realism-NSFW-LoRA-v1"
)
LORA_WEIGHT_NAME = "Japanese-Female-Anatomy.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.75


scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

gc.collect()
if torch.cuda.is_available():
    torch.cuda.empty_cache()

scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    scheduler=scheduler,
    torch_dtype=DTYPE,
    low_cpu_mem_usage=True,
)

pipe.load_lora_weights(
    LIGHTNING_MODEL_ID,
    weight_name=LIGHTNING_WEIGHT_NAME,
    adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="japanese_anatomy_v1",
)
pipe.set_adapters(["japanese_anatomy_v1"], adapter_weights=[LORA_SCALE])

pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Japanese woman, coherent "
    "external female anatomy, natural pelvic structure, unretouched skin "
    "texture, physically plausible pose, soft natural light"
)
negative_prompt = (
    "child, minor, underage, youthful, low quality, blurry, distorted anatomy, "
    "extra limbs, duplicated body parts, fused anatomy, watermark, logo, text"
)

generator = torch.Generator(device="cpu").manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=768,
    height=1024,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("japanese_female_anatomy_realism_v1_low_vram.png")

</details>

🔞 Responsible Use

This model is intended exclusively for lawful, consensual, adult creative work. Do not use it to:

—create or sexualize minors or youthful-looking persons
—create non-consensual or exploitative sexual content
—depict an identifiable real person without permission
—harass, threaten, exploit, or defame another person
—create illegal content or bypass platform rules

Users are responsible for ensuring that their prompts, source material, and generated content comply with applicable laws and platform policies.

License & Copyright Notice

This LoRA may be used for personal and non-commercial purposes only. See the repository LICENSE file for the complete terms.

For commercial licensing inquiries or other permissions, please contact the repository owner.