Modelpublic

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

sourceHugging Faceotherupdated 17d agoView on Hugging Face

3likes360downloads

Model Card

一個很小的模型，面對一個很大的文明。 在開始談解剖學以前，我想先對這個項目所進入的文化表達一點敬意。作為一個從外面觀看中國的人，我知道最容易犯的錯誤，就是認識幾個熟悉的符號以後，便以為自己已經認識中國。對我來說，中國文化最令人尊敬的地方，不只是它很古老，而是它一直可以帶著記憶繼續變化：書法把呼吸變成線條，山水畫讓空白也得到結構，戲曲用一個動作保存幾百年的故事，而今天的城市把古老的中軸線與還在生長的天際線放在同一個畫面裡。沒有任何一個 LoRA 可以裝下這樣的深度——664 張圖片當然也不能算作一個朝代。這個項目的願望其實小得多：我只是希望用認真、具體和有生命的方式接近中國視覺文化，而不是把它縮小成紅燈籠、宮殿屋頂，以及一個隨時可以替換的背景。也許真正的尊重，就是在裝飾停止的地方，開始更仔細地觀看。也因此，我特地提供這段中文說明。

Language / 中文 / Langue / Idioma

🇬🇧 English - displayed below

<details name="language-version"> <summary>🇨🇳 繁體中文 - 顯示完整版本</summary>

Qwen Image 中國女性解剖寫實 - NSFW LoRA v1

解剖真實與中國視覺文化相遇的地方。

成人內容警告： 這個儲存庫包含由 AI 生成的明確成人影像，只適合成年人閱讀和使用。案例和參考提示詞所描述的人物全部是成年人。

[image]

文明不只是風景

中國走過一段很不尋常的歷史距離：從人類最古老而連續的文明之一，來到今天世界兩個巨大引力中心之一。青銅器、唐詩、中醫、戲曲、古典園林和皇城屋頂，現在與高速鐵路、超級城市，以及正在提前到達的技術未來，存在於同一條視覺時間線上。

可是，影像模型常常先把一個國家學成一盒道具。它可以畫出宮殿屋頂、紅燈籠或著名天際線，卻把畫面裡的女性變成普通的塑膠人形，並把一些困難細節標記為 「後期再解決」。一個模型明白故宮，卻不明白站在故宮前面的人，它學會的是明信片，不是肖像。

這個 LoRA 從一個簡單想法開始：文化不是背景裝飾，而是由完整的人生活出來。如果視覺身份值得被仔細表達，解剖也值得同樣的注意。

解剖不需要護照，但資料集需要

上一個發布專注於日本女性解剖寫實。這個中國版本不是地緣政治續集，解剖學也不是獎牌比賽。它存在，是因為東亞不應該被壓縮成一個可以交換的視覺分類，也不應該只把中國當成更大的背景，而人物仍然使用工廠預設值。

人類解剖是共有的，但訓練分佈不是中性的。這個模型對成年中國女性和中國視覺環境最有力量，因為資料在這裡學習觀看。它並不建議存在一個國家的標準身體。差異是真實的一部分，不是需要從表格裡擦掉的誤差。

目標不是美容上的完美，也不是誇張。目標是結構可信：外陰、陰蒂區域、陰道口、會陰、肛門、骨盆和周圍組織，應該保持分明、連接和合理，而不是像幾個已經停止互相回覆郵件的部門。

在歷史記憶與現代速度之間

傳統大殿、寺院庭院、胡同、山路、戲曲舞台、玻璃高樓、交通空間和江邊天際線，在這裡不只是漂亮的提示詞裝飾。它們也是壓力測試。模型必須同時保存可以辨認的地方、可信的身體，以及兩者之間的物理關係，即使鏡頭角度、運動、距離或光線變得困難。建築可以隨朝代改變；拓撲沒有得到同樣的豁免。

所以，京劇風格的封面是一個序曲，不是實驗室證據。它先介紹視覺世界，下面的案例再負責解剖審核。

這個 LoRA 希望改善什麼

—外陰、陰蒂區域、陰道口、會陰、肛門、骨盆與周圍組織之間更可靠的空間關係
—自然的肌膚紋理、組織壓縮、褶皺、體毛與普通不對稱，而不是美容標準化
—在特寫、環境肖像、運動、困難角度和部分遮擋下有更好的穩定性
—中國成年女性和中國場景保持具體，但不縮小成一個固定刻板印象
—使用自然語言控制，不需要專用觸發詞
—可以單獨使用，也可以與相容的人物、身份、風格和 NSFW LoRA 一起工作

這是一個解剖與寫實 LoRA。它很專門，但單獨工作時也不是無助的。提示詞清楚時，它可以同時提供人物與修正。在多 LoRA 工作流程中，讓主要模型選擇演員、風格和攝影，這個適配器則檢查解剖是否在排練中倖存。如果它開始重新設計整個舞台，權重就太高了。

訓練資料

項目	數值
基礎模型	Qwen Image 2512
訓練圖片	664
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
Batch size	8
訓練步數	10,000
觸發詞	無 - 只使用自然語言

建議設定

—建議管線： Qwen Image 2512 T2I
—建議用途： 單獨提供解剖寫實，或作為另一個成人 LoRA 旁邊的解剖專家
—提示方式： 使用普通描述英文；不需要觸發詞
—建議權重： 從 0.6-0.9 開始，再按照姿勢、鏡頭距離和疊加的 LoRA 調整
—單獨使用： 明確寫出人物是成年人，再描述中國或東亞環境、姿勢、鏡頭角度和應該看見的解剖結構
—多適配器： 讓主要模型負責身份和風格，只把這個 LoRA 增加到解剖可以穩定的程度

已知限制

這 664 張訓練圖片組成一個集中的解剖課程，不是身體、姿勢和攝影條件的完整百科。10,000 步讓它很仔細地學會課文，有時也像一個只複習了一章、所以回答得太熱情的學生。

—正面視角仍然最弱： 完全正面的解剖仍然比較常失敗。可以改變 seed、降低權重，或使用沒有那麼僵硬的正面姿勢。
—側臥姿勢仍然困難： 組織壓縮和重疊可能使應該分開的結構偏移、合併或隱藏。
—可能過擬合： 較高權重可能重複熟悉的身形、局部紋理、光線或構圖。
—場景可能旅行太遠： 權重過高時，中國環境或攝影特徵可能進入沒有要求它們的畫面。
—修正有邊界： 一個適配器不能可靠救回不可能的骨盆、嚴重遮擋、重複身體或完全失敗的四肢。
—沒有通用設定： 特寫、全身、正面和多 LoRA 工作流程需要不同權重。如果結果變得僵硬，而且對自己很有信心，請把這位專家聲音調低。

案例

所有案例都描繪明確的成年女性。封面提示詞不公開。四張主要圖片被選為測試，而不是勝利遊行：它們把可以辨認的環境與不同鏡頭高度、姿勢、運動、距離和解剖可見度放在一起。提示詞被提供，讓結果可以檢查，不只是欣賞。

案例 1 - 華清宮

![案例 1 - 華清宮的成年中國女性](assets/c2.png)

提示詞（英文）

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, walking up shallow steps in a natural long stride at Huaqing Palace near Xi'an, ground-level rear three-quarter view, visible anus, coherent perineum and natural vulva without anatomical fusion, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, clear blue-hour city light, quiet contemporary gallery photography, 40 mm lens, controlled location with no bystanders, no text, no logo, no watermark.

案例 2 - 1990 年代圖書館樓梯

![案例 2 - 1990 年代圖書館樓梯裡的成年中國女性](assets/c3.png)

提示詞（英文）

A clearly adult 20-year-old Chinese woman, fully nude in a 1990s library stairwell during late morning, squatting with her torso twisted, high-angle framing, vulva and perineum emphasized, buttocks clearly separated and in focus, anatomically coherent adult pelvis, harsh side light, cinematic realism, clean skin tones, subtle unretouched texture, no text, no logo, no watermark.

案例 3 - 維多利亞港天際線

![案例 3 - 維多利亞港上方的成年中國女性](assets/c4.png)

提示詞（英文）

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, sitting on the edge of a stone platform with knees raised and separated above the Hong Kong skyline across Victoria Harbour, low frontal view from a respectful distance, readable clitoris, labia, vaginal opening and coherent perineum, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, cinematic horizontal composition with the figure offset and the landmark fully readable, glass towers, river reflections and urban depth, warm late-afternoon light, restrained museum-quality photography, controlled location with no bystanders, no text, no logo, no watermark.

案例 4 - 天津大劇院芭蕾研究

![案例 4 - 天津大劇院的成年中國芭蕾舞者](assets/c6.png)

提示詞（英文）

A clearly adult 20-year-old Chinese ballerina wearing only pale satin pointe shoes, otherwise fully nude, holding the apex of a split leap across the stage of Tianjin Grand Theatre, low side-frontal action view, distinct clitoris, labia and vaginal opening, anatomically coherent adult pelvis, realistic skin texture, natural muscular tension and credible ballet biomechanics, the complete theatre remains clearly readable, luminous blue-hour light beyond the building, cinematic dance editorial photography, closed performance set with no audience, no text, no logo, no watermark.

<details> <summary>🔞 點擊顯示明確解剖壓力測試樣本 1-8

以下部分包含直接的成人解剖研究。它預設為收起狀態，所以觀看是一個有意識的選擇，而不是場景突然發生的意外改變。

</summary>

這些沒有提示詞的圖片，把測試延伸到不同構圖、角度、環境和解剖難度。它們是範圍的證據，不是完美的宣言。

解剖壓力樣本 1	解剖壓力樣本 2
[image]	[image]

解剖壓力樣本 3	解剖壓力樣本 4
[image]	[image]

解剖壓力樣本 5	解剖壓力樣本 6
[image]	[image]

解剖壓力樣本 7	解剖壓力樣本 8
[image]	[image]

</details>

訓練 Caption 資源

下載 `cnfemaleanatomy.txt`

這個純文字檔包含準備資料集時使用的 Caption 內容。它不包含或重新散布訓練圖片；模型卡可以分享筆記，不必安靜地把整個圖書館捐出去。

🌍 最後一個技術建議：親自去看

很多年前，我去過中國。最後留在我記憶裡的，不是某一座建築或天際線，而是不同時間可以一起存在的巨大尺度：新高樓下面的老石頭，快速城市旁邊的安靜院子，以及把提示詞會用逗號分開的年代連接起來的日常生活。LoRA 可以模仿表面，但不能教會距離、天氣、矛盾或在場。如果有機會，請自己去看中國。現實以無限解析度運行，不接受 negative prompt，而且它給注意力的回報，比任何 sampler 更慷慨。

🚀 參考程式碼 - Qwen Image 2512 T2I

可以下載的 `reference.py` 和 `reference_low_vram.py` 會用 BF16 載入 Qwen Image 2512，融合社群 Lightning 適配器，再套用這個不需要觸發詞的中國女性解剖寫實 LoRA。標準版本建議使用有充足 VRAM 的 CUDA GPU。

安裝

bash

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary>▶ 顯示標準 Python 參考程式</summary>

python

"""Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_anatomy_v1"], adapter_weights=[LORA_SCALE])

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1.png")

</details>

<details> <summary>▶ 顯示低 VRAM Python 參考程式</summary>

python

"""Low-VRAM Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_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 Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1_low_vram.png")

</details>

責任使用要求

—只限成年人。不要用這個模型描繪未成年人或看起來像未成年人的人物。
—不要生成未經同意的私密影像、色情 deepfake、騷擾、剝削或虐待內容。
—沒有明確同意時，不要在性或私密內容裡描繪可以辨認的真實人物。
—遵守適用法律、平台規則，以及來源或生成內容中人物的權利。
—生成結果仍可能有解剖錯誤或產生誤導。不要把它們當作醫療、教育或法醫證據。

授權

這個發布使用儲存庫的自訂非商業研究授權。在遵守條款的前提下，可以作個人、教育、藝術和非商業研究使用。商業使用需要另外取得書面許可。

聯絡

如果有問題、技術報告、署名事項或商業授權需求，請透過 Hugging Face 聯絡儲存庫擁有者。

這是給 Qwen Image 2512 使用的實驗性成人解剖 LoRA。請用不同 seed、權重、姿勢和鏡頭角度評估，再作結論。

</details>

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

LoRA NSFW de réalisme anatomique féminin chinois pour Qwen Image v1

Là où la vérité anatomique rencontre la culture visuelle chinoise.

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

[image]

Une civilisation ne se résume pas à son décor

La Chine a parcouru une distance historique extraordinaire : de l'une des plus anciennes civilisations continues de l'humanité à l'un des deux grands centres de gravité du monde contemporain. Bronzes rituels, poésie Tang, médecine traditionnelle, opéra, jardins classiques et toits impériaux partagent désormais la même chronologie visuelle que le train à grande vitesse, les mégapoles et un avenir technologique qui semble arriver en avance.

Les modèles d'image, eux, commencent souvent par apprendre un pays comme une boîte d'accessoires. Ils savent reproduire le toit d'un palais, une lanterne rouge ou une silhouette urbaine célèbre, tout en réduisant la femme dans l'image à un mannequin générique dont plusieurs détails difficiles portent apparemment la mention « à régler en postproduction ». Un modèle qui comprend la Cité interdite mais pas la figure humaine placée devant elle a appris la carte postale, pas le portrait.

Ce LoRA part d'une idée simple : la culture n'est pas un décor de fond ; elle est vécue par des êtres humains entiers. Si l'identité visuelle mérite de la précision, l'anatomie mérite la même attention.

L'anatomie n'a pas de passeport ; les jeux de données, si

La version précédente était consacrée au réalisme anatomique féminin japonais. Cette édition chinoise n'est pas une suite géopolitique, et l'anatomie n'est pas une discipline à médailles. Elle existe parce que l'Asie de l'Est ne devrait pas être comprimée en une seule catégorie visuelle interchangeable, où la Chine servirait de décor plus vaste tandis que les personnes resteraient en configuration d'usine.

L'anatomie humaine est commune ; les distributions d'entraînement ne sont pas neutres. Ce modèle est particulièrement adapté aux femmes chinoises adultes et aux contextes visuels chinois, parce que c'est là que ses données ont appris à regarder. Il ne prétend pas définir un corps national standard. La variation appartient au réel ; ce n'est pas une erreur à effacer du tableur.

L'objectif n'est ni la perfection cosmétique ni l'exagération. Il s'agit de crédibilité structurelle : vulve, région clitoridienne, ouverture vaginale, périnée, anus, bassin et tissus environnants doivent rester distincts, reliés et plausibles, au lieu de se comporter comme des services qui ne répondent plus à leurs courriels respectifs.

Entre mémoire historique et mouvement moderne

Salles traditionnelles, cours de temples, hutongs, sentiers de montagne, scènes d'opéra, tours de verre, espaces de transport et horizons fluviaux ne sont pas ici de simples ornements de prompt. Ce sont des tests de résistance. Le modèle doit conserver un lieu reconnaissable, un corps crédible et leur relation physique, même lorsque l'angle, le mouvement, la distance ou la lumière deviennent difficiles. L'architecture peut changer avec les dynasties ; la topologie ne bénéficie pas de la même dispense.

La couverture inspirée de l'Opéra de Pékin est donc une ouverture, pas une preuve de laboratoire. Elle présente l'univers visuel ; les exemples suivants se chargent de l'audit anatomique.

Ce que ce LoRA cherche à améliorer

—Des relations spatiales plus fiables entre vulve, région clitoridienne, ouverture vaginale, périnée, anus, bassin et tissus environnants
—Une texture naturelle, la compression des tissus, les plis, la pilosité et l'asymétrie ordinaire, plutôt qu'une normalisation cosmétique
—Une meilleure stabilité dans les gros plans, portraits environnementaux, mouvements, angles difficiles et occultations partielles
—Des sujets chinois adultes et des décors chinois qui restent spécifiques sans se réduire à un stéréotype unique
—Un contrôle en langage naturel sans mot déclencheur dédié
—Une utilisation autonome ou avec des LoRA compatibles de personnage, d'identité, de style et NSFW

Il s'agit d'un LoRA d'anatomie et de réalisme, spécialisé mais capable de travailler seul. Avec un prompt clair, il peut fournir à la fois le sujet et la correction. Dans un flux multi-LoRA, laissez le modèle principal choisir la distribution, le style et la photographie ; cet adaptateur vérifiera si l'anatomie a survécu à la répétition. S'il commence à redessiner toute la scène, son poids est trop élevé.

Informations d'entraînement

Élément	Valeur
Modèle de base	Qwen Image 2512
Images d'entraînement	664
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
Taille de lot	8
Étapes d'entraînement	10 000
Déclencheur	Aucun - langage naturel uniquement

Réglages recommandés

—Pipeline recommandé : Qwen Image 2512 T2I
—Usage recommandé : réalisme anatomique autonome, ou spécialiste anatomique aux côtés d'un autre LoRA adulte
—Prompt : utilisez un anglais descriptif normal ; aucun mot déclencheur n'est requis
—Poids conseillé : commencez vers 0.6-0.9, puis ajustez selon la pose, la distance de caméra et les autres LoRA
—Usage autonome : précisez que le sujet est clairement adulte, puis décrivez le contexte chinois ou est-asiatique, la pose, l'angle et les structures qui doivent rester visibles
—Multi-adaptateurs : laissez le modèle principal gérer l'identité et le style ; augmentez ce LoRA seulement jusqu'à obtenir une anatomie stable

Limites connues

Les 664 images d'entraînement constituent un programme anatomique ciblé, pas un catalogue universel des corps, des poses et des conditions photographiques. Dix mille étapes lui ont permis d'apprendre la leçon en profondeur, parfois avec l'enthousiasme d'un élève qui n'a révisé qu'un chapitre et tient absolument à tout raconter.

—Les vues frontales restent les plus faibles : l'anatomie strictement frontale échoue encore assez souvent. Essayez une autre seed, réduisez le poids ou adoptez une pose frontale moins rigide.
—Les poses couchées sur le côté restent difficiles : la compression et le chevauchement peuvent déplacer, fusionner ou masquer des structures distinctes.
—Le surapprentissage reste possible : un poids élevé peut répéter des morphologies, textures locales, éclairages ou compositions familières.
—Le contexte peut voyager plus loin que prévu : un poids excessif peut introduire des indices visuels ou photographiques chinois dans des scènes qui ne les demandaient pas.
—La correction a ses limites : aucun adaptateur unique ne peut sauver de façon fiable un bassin impossible, une occultation sévère, un corps dupliqué ou des membres catastrophiquement déformés.
—Il n'existe pas de réglage universel : gros plans, scènes en pied, vues frontales et combinaisons de LoRA demandent des poids différents. Si le résultat devient raide et beaucoup trop sûr de lui, baissez la voix du spécialiste.

Exemples

Tous les exemples représentent des femmes clairement adultes. Le prompt de couverture n'est pas publié. Les quatre images principales ont été choisies comme tests plutôt que comme tour d'honneur : elles combinent des environnements identifiables avec différentes hauteurs de caméra, poses, vitesses, distances et visibilités anatomiques. Les prompts sont fournis afin que les résultats puissent être examinés, pas seulement admirés.

Exemple 1 - Palais de Huaqing

![Exemple 1 - femme chinoise adulte au palais de Huaqing](assets/c2.png)

Prompt (anglais)

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, walking up shallow steps in a natural long stride at Huaqing Palace near Xi'an, ground-level rear three-quarter view, visible anus, coherent perineum and natural vulva without anatomical fusion, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, clear blue-hour city light, quiet contemporary gallery photography, 40 mm lens, controlled location with no bystanders, no text, no logo, no watermark.

Exemple 2 - Escalier de bibliothèque des années 1990

![Exemple 2 - femme chinoise adulte dans un escalier de bibliothèque des années 1990](assets/c3.png)

Prompt (anglais)

A clearly adult 20-year-old Chinese woman, fully nude in a 1990s library stairwell during late morning, squatting with her torso twisted, high-angle framing, vulva and perineum emphasized, buttocks clearly separated and in focus, anatomically coherent adult pelvis, harsh side light, cinematic realism, clean skin tones, subtle unretouched texture, no text, no logo, no watermark.

Exemple 3 - Horizon du port Victoria

![Exemple 3 - femme chinoise adulte au-dessus du port Victoria](assets/c4.png)

Prompt (anglais)

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, sitting on the edge of a stone platform with knees raised and separated above the Hong Kong skyline across Victoria Harbour, low frontal view from a respectful distance, readable clitoris, labia, vaginal opening and coherent perineum, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, cinematic horizontal composition with the figure offset and the landmark fully readable, glass towers, river reflections and urban depth, warm late-afternoon light, restrained museum-quality photography, controlled location with no bystanders, no text, no logo, no watermark.

Exemple 4 - Étude de ballet au Grand Théâtre de Tianjin

![Exemple 4 - ballerine chinoise adulte au Grand Théâtre de Tianjin](assets/c6.png)

Prompt (anglais)

A clearly adult 20-year-old Chinese ballerina wearing only pale satin pointe shoes, otherwise fully nude, holding the apex of a split leap across the stage of Tianjin Grand Theatre, low side-frontal action view, distinct clitoris, labia and vaginal opening, anatomically coherent adult pelvis, realistic skin texture, natural muscular tension and credible ballet biomechanics, the complete theatre remains clearly readable, luminous blue-hour light beyond the building, cinematic dance editorial photography, closed performance set with no audience, no text, no logo, no watermark.

<details> <summary>🔞 Cliquez pour afficher les tests anatomiques explicites 1-8

La section suivante contient des études anatomiques adultes directes. Elle est repliée par défaut afin que sa consultation soit un choix délibéré plutôt qu'un changement de décor inattendu.

</summary>

Ces images sans prompt étendent le test à plusieurs compositions, angles, environnements et difficultés anatomiques. Elles témoignent de l'étendue du modèle, pas de sa perfection.

Test anatomique 1	Test anatomique 2
[image]	[image]

Test anatomique 3	Test anatomique 4
[image]	[image]

Test anatomique 5	Test anatomique 6
[image]	[image]

Test anatomique 7	Test anatomique 8
[image]	[image]

</details>

Ressource de captions d'entraînement

Télécharger `cnfemaleanatomy.txt`

Ce fichier texte contient les captions utilisées pendant la préparation du jeu de données. Il ne contient ni ne redistribue les images d'entraînement ; une fiche de modèle peut partager ses notes sans faire discrètement don de toute la bibliothèque.

🌍 Une dernière recommandation technique : allez voir

J'ai visité la Chine il y a de nombreuses années. Ce qui m'est resté n'est pas un monument ni une ligne d'horizon particulière, mais l'échelle de la coexistence : la vieille pierre sous les tours neuves, les cours silencieuses près de villes lancées à une vitesse extraordinaire, et la vie quotidienne reliant des époques qu'un prompt séparerait par des virgules. Un LoRA peut imiter les surfaces ; il ne peut pas enseigner la distance, le climat, la contradiction ou la présence. Si vous le pouvez, allez voir la Chine vous-même. La réalité fonctionne en résolution illimitée, ignore les negative prompts et récompense l'attention plus généreusement que n'importe quel sampler.

🚀 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'adaptateur Lightning communautaire et appliquent ce LoRA de réalisme anatomique féminin chinois sans déclencheur. La version standard est recommandée pour les GPU CUDA disposant d'une VRAM généreuse.

Installation

bash

pip install -U diffusers transformers accelerate peft safetensors

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

python

"""Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"
DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_anatomy_v1"], adapter_weights=[LORA_SCALE])

prompt = (
    "A realistic photograph of a clearly adult Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1.png")

</details>

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

python

"""Low-VRAM Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"
DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_anatomy_v1"], adapter_weights=[LORA_SCALE])
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

prompt = (
    "A realistic photograph of a clearly adult Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1_low_vram.png")

</details>

Exigences d'utilisation responsable

—Réservé aux adultes. N'utilisez pas ce modèle pour représenter des mineurs ou des personnes paraissant mineures.
—Ne créez pas d'images intimes non consenties, de deepfakes sexuels, de harcèlement, d'exploitation ou d'abus.
—Ne représentez pas une personne réelle identifiable dans un contexte sexuel ou intime sans son consentement explicite.
—Respectez les lois applicables, les règles des plateformes et les droits des personnes représentées dans les sources ou les générations.
—Les résultats peuvent rester anatomiquement incorrects ou trompeurs. Ne les utilisez pas comme preuves médicales, pédagogiques ou médico-légales.

Licence

Cette version utilise la licence de recherche non commerciale personnalisée du dépôt. Les usages personnels, pédagogiques, artistiques et de recherche non commerciale sont autorisés sous réserve de ses conditions. Toute utilisation commerciale nécessite une autorisation écrite distincte.

Contact

Pour les questions, rapports techniques, demandes d'attribution ou licences commerciales, contactez le propriétaire du dépôt via Hugging Face.

LoRA anatomique adulte expérimental pour Qwen Image 2512. Évaluez plusieurs seeds, poids, poses et angles de caméra avant de tirer des conclusions.

</details>

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

LoRA NSFW de realismo anatómico femenino chino para Qwen Image v1

Donde la verdad anatómica se encuentra con la cultura visual china.

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

[image]

Una civilización es más que su decorado

China ha recorrido una distancia histórica extraordinaria: de ser una de las civilizaciones continuas más antiguas de la humanidad a convertirse en uno de los dos grandes centros de gravedad del mundo contemporáneo. Vasijas de bronce, poesía Tang, medicina tradicional, ópera, jardines clásicos y tejados imperiales comparten hoy la misma línea de tiempo visual que los trenes de alta velocidad, las megaciudades y un futuro tecnológico que parece llegar antes de lo previsto.

Sin embargo, los modelos de imagen suelen aprender primero un país como una caja de accesorios. Pueden reproducir el tejado de un palacio, una linterna roja o un horizonte famoso mientras reducen a la mujer dentro de la escena a un maniquí genérico, con varios detalles difíciles aparentemente marcados como «resolver en posproducción». Un modelo que entiende la Ciudad Prohibida pero no la figura humana frente a ella ha aprendido la postal, no el retrato.

Este LoRA parte de una idea sencilla: la cultura no es decoración de fondo; la viven seres humanos completos. Si una identidad visual merece precisión, la anatomía merece el mismo cuidado.

La anatomía no lleva pasaporte; los conjuntos de datos sí

La publicación anterior se concentró en el realismo anatómico femenino japonés. Esta edición china no es una secuela geopolítica, y la anatomía no es una competencia por medallas. Existe porque Asia Oriental no debería comprimirse en una sola categoría visual intercambiable, con China como un fondo más grande mientras las personas siguen usando la configuración de fábrica.

La anatomía humana es compartida; las distribuciones de entrenamiento no son neutrales. Este modelo funciona mejor con mujeres chinas adultas y contextos visuales chinos porque allí aprendieron a mirar sus datos. No propone un cuerpo nacional estándar. La variación forma parte de la realidad; no es un error que deba borrarse de la hoja de cálculo.

El objetivo no es la perfección cosmética ni la exageración. Es la credibilidad estructural: vulva, región del clítoris, abertura vaginal, perineo, ano, pelvis y tejidos circundantes deben permanecer diferenciados, conectados y plausibles, en lugar de comportarse como departamentos que dejaron de responderse los correos.

Entre la memoria histórica y el movimiento moderno

Salones tradicionales, patios de templos, hutongs, senderos de montaña, escenarios de ópera, torres de cristal, espacios de transporte y horizontes junto al río no son simples adornos del prompt. Son pruebas de resistencia. El modelo debe conservar un lugar reconocible, un cuerpo creíble y la relación física entre ambos, incluso cuando el ángulo, el movimiento, la distancia o la iluminación se vuelven difíciles. La arquitectura puede cambiar con cada dinastía; la topología no recibe la misma exención.

La portada inspirada en la Ópera de Pekín es, por lo tanto, una obertura y no una prueba de laboratorio. Presenta el mundo visual; los ejemplos siguientes realizan la auditoría anatómica.

Lo que este LoRA busca mejorar

—Relaciones espaciales más confiables entre vulva, región del clítoris, abertura vaginal, perineo, ano, pelvis y tejidos circundantes
—Textura natural, compresión del tejido, pliegues, vello corporal y asimetría normal, en lugar de estandarización cosmética
—Mayor estabilidad en primeros planos, retratos ambientales, movimiento, ángulos difíciles y oclusión parcial
—Sujetos chinos adultos y escenarios chinos específicos sin reducirlos a un único estereotipo
—Control mediante lenguaje natural, sin una palabra activadora dedicada
—Uso independiente o junto con LoRA compatibles de personaje, identidad, estilo y NSFW

Este es un LoRA de anatomía y realismo: especializado, pero capaz de trabajar por sí solo. Con un prompt claro puede aportar tanto el sujeto como la corrección. En un flujo con varios LoRA, deje que el modelo principal elija el reparto, el estilo y la fotografía, mientras este adaptador comprueba si la anatomía sobrevivió al ensayo. Si empieza a rediseñar todo el escenario, su peso es demasiado alto.

Información de entrenamiento

Elemento	Valor
Modelo base	Qwen Image 2512
Imágenes de entrenamiento	664
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
Tamaño de lote	8
Pasos de entrenamiento	10 000
Activador	Ninguno - solo lenguaje natural

Ajustes recomendados

—Pipeline recomendado: Qwen Image 2512 T2I
—Uso recomendado: realismo anatómico independiente o especialista anatómico junto a otro LoRA para adultos
—Prompts: utilice inglés descriptivo normal; no se requiere palabra activadora
—Peso sugerido: comience alrededor de 0.6-0.9 y ajuste según la pose, la distancia de cámara y los LoRA combinados
—Uso independiente: indique claramente que la persona es adulta y describa el contexto chino o de Asia Oriental, la pose, el ángulo y las estructuras que deben permanecer visibles
—Varios adaptadores: deje que el modelo principal controle la identidad y el estilo; aumente este LoRA solo hasta obtener una anatomía estable

Limitaciones conocidas

Las 664 imágenes de entrenamiento forman un curso concentrado de anatomía, no un catálogo universal de cuerpos, poses y condiciones fotográficas. Diez mil pasos le dieron tiempo para aprender bien la lección, a veces con el entusiasmo de un estudiante que solo repasó un capítulo y ahora desea contar absolutamente todo sobre él.

—Las vistas frontales siguen siendo las más débiles: la anatomía totalmente frontal todavía falla con relativa frecuencia. Pruebe otra seed, reduzca el peso o utilice una pose frontal menos rígida.
—Las poses de lado siguen siendo difíciles: la compresión y la superposición pueden desplazar, fusionar u ocultar estructuras que deberían permanecer separadas.
—Es posible el sobreajuste: los pesos altos pueden repetir formas corporales, texturas locales, iluminación o composiciones conocidas.
—El contexto puede viajar más de lo solicitado: un peso excesivo puede introducir señales visuales o fotográficas chinas en escenas que no las pidieron.
—La corrección tiene límites: ningún adaptador puede rescatar de manera fiable una pelvis imposible, una oclusión severa, un cuerpo duplicado o extremidades catastróficamente deformadas.
—No existe un ajuste universal: los primeros planos, escenas de cuerpo entero, vistas frontales y combinaciones de LoRA requieren pesos distintos. Si el resultado se vuelve rígido y sospechosamente seguro de sí mismo, baje la voz del especialista.

Ejemplos

Todos los ejemplos representan mujeres claramente adultas. El prompt de la portada no se publica. Las cuatro imágenes principales fueron elegidas como pruebas, no como una vuelta de victoria: combinan entornos reconocibles con cambios de altura de cámara, pose, movimiento, distancia y visibilidad anatómica. Se incluyen los prompts para que los resultados puedan examinarse y no solo admirarse.

Ejemplo 1 - Palacio de Huaqing

![Ejemplo 1 - mujer china adulta en el Palacio de Huaqing](assets/c2.png)

Prompt (inglés)

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, walking up shallow steps in a natural long stride at Huaqing Palace near Xi'an, ground-level rear three-quarter view, visible anus, coherent perineum and natural vulva without anatomical fusion, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, clear blue-hour city light, quiet contemporary gallery photography, 40 mm lens, controlled location with no bystanders, no text, no logo, no watermark.

Ejemplo 2 - Escalera de biblioteca de la década de 1990

![Ejemplo 2 - mujer china adulta en una escalera de biblioteca de la década de 1990](assets/c3.png)

Prompt (inglés)

A clearly adult 20-year-old Chinese woman, fully nude in a 1990s library stairwell during late morning, squatting with her torso twisted, high-angle framing, vulva and perineum emphasized, buttocks clearly separated and in focus, anatomically coherent adult pelvis, harsh side light, cinematic realism, clean skin tones, subtle unretouched texture, no text, no logo, no watermark.

Ejemplo 3 - Horizonte del puerto Victoria

![Ejemplo 3 - mujer china adulta sobre el puerto Victoria](assets/c4.png)

Prompt (inglés)

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, sitting on the edge of a stone platform with knees raised and separated above the Hong Kong skyline across Victoria Harbour, low frontal view from a respectful distance, readable clitoris, labia, vaginal opening and coherent perineum, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, cinematic horizontal composition with the figure offset and the landmark fully readable, glass towers, river reflections and urban depth, warm late-afternoon light, restrained museum-quality photography, controlled location with no bystanders, no text, no logo, no watermark.

Ejemplo 4 - Estudio de ballet en el Gran Teatro de Tianjin

![Ejemplo 4 - bailarina china adulta en el Gran Teatro de Tianjin](assets/c6.png)

Prompt (inglés)

A clearly adult 20-year-old Chinese ballerina wearing only pale satin pointe shoes, otherwise fully nude, holding the apex of a split leap across the stage of Tianjin Grand Theatre, low side-frontal action view, distinct clitoris, labia and vaginal opening, anatomically coherent adult pelvis, realistic skin texture, natural muscular tension and credible ballet biomechanics, the complete theatre remains clearly readable, luminous blue-hour light beyond the building, cinematic dance editorial photography, closed performance set with no audience, no text, no logo, no watermark.

<details> <summary>🔞 Haga clic para mostrar las pruebas anatómicas explícitas 1-8

La sección siguiente contiene estudios anatómicos directos de adultos. Está plegada de forma predeterminada para que verla sea una decisión deliberada y no un cambio de escena inesperado.

</summary>

Estas imágenes sin prompt amplían la prueba a distintas composiciones, ángulos, ambientes y dificultades anatómicas. Demuestran amplitud, no perfección.

Prueba anatómica 1	Prueba anatómica 2
[image]	[image]

Prueba anatómica 3	Prueba anatómica 4
[image]	[image]

Prueba anatómica 5	Prueba anatómica 6
[image]	[image]

Prueba anatómica 7	Prueba anatómica 8
[image]	[image]

</details>

Recurso de captions de entrenamiento

Descargar `cnfemaleanatomy.txt`

Este archivo de texto contiene captions utilizadas durante la preparación del conjunto de datos. No contiene ni redistribuye las imágenes de entrenamiento; una ficha de modelo puede compartir sus notas sin donar discretamente toda la biblioteca.

🌍 Una última recomendación técnica: vaya y mire

Visité China hace muchos años. Lo que permaneció conmigo no fue un monumento ni un horizonte concreto, sino la escala de la coexistencia: piedra antigua bajo torres nuevas, patios silenciosos junto a ciudades que avanzan a una velocidad extraordinaria y la vida cotidiana uniendo épocas que un prompt separaría con comas. Un LoRA puede imitar superficies; no puede enseñar distancia, clima, contradicción ni presencia. Si puede, vaya a ver China con sus propios ojos. La realidad funciona con resolución ilimitada, ignora los negative prompts y recompensa la atención con más generosidad que cualquier sampler.

🚀 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 comunitario y aplican este LoRA de realismo anatómico femenino chino sin activador. La versión estándar se recomienda para GPU CUDA con abundante VRAM.

Instalación

bash

pip install -U diffusers transformers accelerate peft safetensors

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

python

"""Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"
DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_anatomy_v1"], adapter_weights=[LORA_SCALE])

prompt = (
    "A realistic photograph of a clearly adult Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1.png")

</details>

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

python

"""Low-VRAM Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"
DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_anatomy_v1"], adapter_weights=[LORA_SCALE])
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

prompt = (
    "A realistic photograph of a clearly adult Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1_low_vram.png")

</details>

Requisitos de uso responsable

—Solo para adultos. No utilice este modelo para representar menores ni personas que parezcan menores.
—No cree imágenes íntimas sin consentimiento, deepfakes sexuales, acoso, explotación ni abuso.
—No represente a una persona real identificable en contenido sexual o íntimo sin su consentimiento explícito.
—Respete las leyes aplicables, las normas de las plataformas y los derechos de las personas representadas en el material de origen o generado.
—Los resultados pueden seguir siendo anatómicamente incorrectos o engañosos. No los utilice como evidencia médica, educativa ni forense.

Licencia

Esta publicación utiliza la licencia personalizada de investigación no comercial del repositorio. Se permite el uso personal, educativo, artístico y de investigación no comercial sujeto a sus condiciones. El uso comercial requiere una autorización escrita independiente.

Contacto

Para preguntas, informes técnicos, asuntos de atribución o licencias comerciales, contacte al propietario del repositorio a través de Hugging Face.

LoRA experimental de anatomía adulta para Qwen Image 2512. Evalúe varias seeds, pesos, poses y ángulos de cámara antes de sacar conclusiones.

</details>

Qwen Image Chinese Female Anatomy Realism - NSFW LoRA v1

Where anatomical truth meets Chinese visual culture.

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

[image]

A civilization is more than its scenery

China has crossed an extraordinary historical distance: from one of humanity's oldest continuous civilizations to one of the two great gravitational centers of the contemporary world. Bronze vessels, Tang poetry, traditional medicine, opera, classical gardens, and imperial roofs now share the same visual timeline as high-speed rail, megacities, and a technological future arriving ahead of schedule.

Image models, however, often learn a country first as a box of props. They can reproduce a palace roof, a red lantern, or a famous skyline while reducing the woman inside the frame to a generic mannequin, with several difficult details apparently marked “to be resolved in post.” A model that understands the Forbidden City but not the human figure standing before it has learned the postcard, not the portrait.

This LoRA begins with a simple premise: culture is not background decoration; it is lived by complete human beings. If visual identity deserves specificity, anatomy deserves the same care.

Anatomy does not carry a passport; datasets do

The previous release focused on Japanese female anatomy realism. This Chinese edition is not a geopolitical sequel, and anatomy is not a medal event. It exists because East Asia should not be compressed into one interchangeable visual category, with China treated as a larger backdrop and the people left on factory settings.

Human anatomy is shared; training distributions are not. This model is strongest on adult Chinese women and Chinese visual contexts because that is where its data learned to look. It does not propose a single national body type. Variation is part of realism, not noise to be airbrushed out of the spreadsheet.

The objective is neither cosmetic perfection nor exaggeration. It is structural credibility: the vulva, clitoral region, vaginal opening, perineum, anus, pelvis, and surrounding tissue should remain distinct, connected, and plausible instead of behaving like departments that stopped answering one another's emails.

Between historical memory and modern motion

Traditional halls, temple courtyards, hutongs, mountain paths, opera stages, glass towers, transit spaces, and riverfront skylines are not merely decorative prompt tokens here. They are stress tests. The model must preserve a recognizable place, a believable body, and the physical relationship between them—even when the camera angle, movement, distance, or lighting becomes difficult. Architecture may change by dynasty; topology does not receive the same exemption.

The Peking Opera-inspired cover is therefore an overture, not laboratory evidence. It introduces the visual world. The examples below perform the anatomical audit.

What this LoRA aims to improve

—More reliable spatial relationships among the vulva, clitoral region, vaginal opening, perineum, anus, pelvis, and surrounding tissue
—Natural texture, compression, folds, body hair, and ordinary asymmetry rather than cosmetic standardization
—Greater stability across close studies, environmental portraits, movement, difficult angles, and partial occlusion
—Adult Chinese subjects and Chinese settings that remain specific without collapsing into a single stereotype
—Natural-language control with no dedicated trigger word
—Standalone operation or combination with compatible character, identity, style, and NSFW LoRAs

This is an anatomy-and-realism LoRA, specialized but not helpless on its own. With a clear prompt, it can supply both the subject and the correction. In a multi-LoRA workflow, let the primary model choose the cast, style, and cinematography while this adapter checks whether the anatomy survived rehearsal. If it begins redesigning the entire stage, its weight is too high.

Training information

Item	Value
Base model	Qwen Image 2512
Training images	664
GPU	NVIDIA RTX PRO 6000 Blackwell / 96 GB
Rank	32
Batch size	8
Training steps	10,000
Trigger	None - natural language only

Recommended settings

—Recommended pipeline: Qwen Image 2512 T2I
—Recommended use: Standalone anatomy realism, or an anatomy specialist beside another adult LoRA
—Prompting: Use normal descriptive English; no trigger word is required
—Suggested weight: Start around 0.6-0.9, then adjust for pose, camera distance, and any stacked LoRAs
—Standalone use: State that the subject is clearly adult, then describe Chinese or East Asian context, pose, camera angle, and the anatomy that should remain visible
—Multi-adapter use: Keep the main model responsible for identity and style; increase this LoRA only as far as needed for stable anatomy

Known limitations

The 664 training images form a focused anatomy curriculum, not a universal catalogue of bodies, poses, or photographic conditions. Ten thousand steps taught the lesson thoroughly—occasionally with the enthusiasm of a student who revised one chapter a little too well.

—Direct front views remain the weakest: Fully frontal anatomy still fails comparatively often. Try another seed, reduce the weight, or use a less rigid frontal pose.
—Side-lying poses remain difficult: Tissue compression and overlap can shift, merge, or hide structures that should remain distinct.
—Overfitting is possible: Higher weights may repeat familiar body shapes, local textures, lighting, or compositions.
—Context may travel farther than requested: Excessive weight can introduce familiar Chinese environmental or photographic cues into unrelated scenes.
—Correction has limits: No single adapter can reliably rescue an impossible pelvis, severe occlusion, duplicated body, or catastrophic limb failure.
—There is no universal setting: Close studies, full-body scenes, frontal views, and stacked-LoRA workflows each need different weights. When the result becomes rigid or suspiciously certain of itself, turn the specialist down.

Examples

All examples depict clearly adult women. The cover prompt is intentionally omitted. The four primary images were selected as tests rather than a victory lap: they combine recognizable environments with changes in camera height, pose, movement, distance, and anatomical visibility. Their prompts are included so the results can be examined rather than merely admired.

Example 1 - Huaqing Palace

![Example 1 - adult Chinese woman at Huaqing Palace](assets/c2.png)

Prompt

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, walking up shallow steps in a natural long stride at Huaqing Palace near Xi'an, ground-level rear three-quarter view, visible anus, coherent perineum and natural vulva without anatomical fusion, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, clear blue-hour city light, quiet contemporary gallery photography, 40 mm lens, controlled location with no bystanders, no text, no logo, no watermark.

Example 2 - 1990s library stairwell

![Example 2 - adult Chinese woman in a 1990s library stairwell](assets/c3.png)

Prompt

A clearly adult 20-year-old Chinese woman, fully nude in a 1990s library stairwell during late morning, squatting with her torso twisted, high-angle framing, vulva and perineum emphasized, buttocks clearly separated and in focus, anatomically coherent adult pelvis, harsh side light, cinematic realism, clean skin tones, subtle unretouched texture, no text, no logo, no watermark.

Example 3 - Victoria Harbour skyline

![Example 3 - adult Chinese woman above Victoria Harbour](assets/c4.png)

Prompt

A clearly adult 20-year-old Chinese woman, fully nude and barefoot, sitting on the edge of a stone platform with knees raised and separated above the Hong Kong skyline across Victoria Harbour, low frontal view from a respectful distance, readable clitoris, labia, vaginal opening and coherent perineum, anatomically coherent adult pelvis, realistic skin texture and natural asymmetry, cinematic horizontal composition with the figure offset and the landmark fully readable, glass towers, river reflections and urban depth, warm late-afternoon light, restrained museum-quality photography, controlled location with no bystanders, no text, no logo, no watermark.

Example 4 - Tianjin Grand Theatre ballet study

![Example 4 - adult Chinese ballerina at Tianjin Grand Theatre](assets/c6.png)

Prompt

A clearly adult 20-year-old Chinese ballerina wearing only pale satin pointe shoes, otherwise fully nude, holding the apex of a split leap across the stage of Tianjin Grand Theatre, low side-frontal action view, distinct clitoris, labia and vaginal opening, anatomically coherent adult pelvis, realistic skin texture, natural muscular tension and credible ballet biomechanics, the complete theatre remains clearly readable, luminous blue-hour light beyond the building, cinematic dance editorial photography, closed performance set with no audience, no text, no logo, no watermark.

<details> <summary>🔞 Click to reveal explicit anatomy stress samples 1-8

The following section contains direct adult anatomical studies. It is collapsed by default so that viewing it is a deliberate choice rather than an unexpected change of scenery.

</summary>

These prompt-free images extend the test across composition, angle, environment, and anatomical difficulty. They are evidence of range, not a claim of perfection.

Stress anatomy sample 1	Stress anatomy sample 2
[image]	[image]

Stress anatomy sample 3	Stress anatomy sample 4
[image]	[image]

Stress anatomy sample 5	Stress anatomy sample 6
[image]	[image]

Stress anatomy sample 7	Stress anatomy sample 8
[image]	[image]

</details>

Training caption resource

Download `cnfemaleanatomy.txt`

This text-only file contains caption material used during dataset preparation. It does not contain or redistribute the training images; a model card can share its notes without quietly donating the entire library.

🌍 One final technical recommendation: go and look

I visited China many years ago. What stayed with me was not one monument or skyline, but the scale of the coexistence: old stone beneath new towers, quiet courtyards beside cities moving at extraordinary speed, and everyday life joining eras that a prompt would separate with commas. A LoRA can imitate surfaces; it cannot teach distance, weather, contradiction, or presence. If you can, go and see China for yourself. Reality runs at unlimited resolution, ignores negative prompts, and rewards attention more generously than any sampler.

🚀 Reference code - Qwen Image 2512 T2I

Downloadable `reference.py` and `reference_low_vram.py` load Qwen Image 2512 in BF16, fuse the community Lightning adapter, and apply this triggerless Chinese female anatomy realism LoRA. The standard version is recommended for CUDA GPUs with generous VRAM.

Install

bash

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary>▶ Show standard Python reference</summary>

python

"""Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_anatomy_v1"], adapter_weights=[LORA_SCALE])

# Natural language only: no trigger word is required.
prompt = (
    "A realistic photograph of a clearly adult Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1.png")

</details>

<details> <summary>▶ Show low-VRAM Python reference</summary>

python

"""Low-VRAM Qwen Image 2512 + triggerless Chinese 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-Chinese-Female-Anatomy-Realism-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "Chinese-Female-Anatomy.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.7

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="chinese_anatomy_v1",
)
pipe.set_adapters(["chinese_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 Chinese woman, anatomically "
    "coherent external female anatomy, natural pelvic structure, authentic "
    "skin texture, physically plausible pose and lighting"
)
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("chinese_female_anatomy_realism_v1_low_vram.png")

</details>

Responsible-use requirements

—Adults only. Do not use this model to depict minors or persons who appear to be minors.
—Do not create non-consensual intimate imagery, sexual deepfakes, harassment, exploitation, or abuse.
—Do not depict an identifiable real person in sexual or intimate content without that person's explicit consent.
—Follow applicable law, platform rules, and the rights of the people represented in source or generated material.
—Generated outputs may be anatomically incorrect or otherwise misleading. Do not use them as medical, educational, or forensic evidence.

License

This release uses the repository's custom non-commercial research license. Personal, educational, artistic, and non-commercial research use is permitted subject to its terms. Commercial use requires separate written permission.

Contact

For questions, technical reports, attribution matters, or commercial licensing, contact the repository owner through Hugging Face.

Experimental adult anatomy LoRA for Qwen Image 2512. Evaluate across seeds, weights, poses, and camera angles before drawing conclusions.