Chinese-Female-NSFW-X-LoRA-v1

<div style="display:flex; flex-wrap:nowrap; gap:6px; align-items:center; overflow-x:auto;"><a href="#" style="display:inline-flex; align-items:center; height:22px; vertical-align:middle;"><img alt="Visitors" src="https://visitor-badge.laobi.icu/badge?page_id=perpetual3x.Chinese-Female-NSFW-X-LoRA-v1" style="display:block; height:22px; vertical-align:middle;"></a><a href="#" style="display:inline-flex; align-items:center; height:22px; vertical-align:middle;"><img alt="Base Model" src="https://img.shields.io/badge/Base-Qwen%20Image-blue" style="display:block; height:22px; vertical-align:middle;"></a><a href="#" style="display:inline-flex; align-items:center; height:22px; vertical-align:middle;"><img alt="GPU" src="https://img.shields.io/badge/GPU-RTX%20PRO%206000%20Blackwell%2096GB-blue" style="display:block; height:22px; vertical-align:middle;"></a><a href="#" style="display:inline-flex; align-items:center; height:22px; vertical-align:middle;"><img alt="Training" src="https://img.shields.io/badge/Training-100K%20Steps-purple" style="display:block; height:22px; vertical-align:middle;"></a></div>

Language / Langue / Idioma

🇬🇧 English — displayed below

---

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

---

LoRA NSFW X de femmes chinoises v1

Photographies du quotidien, réalité imparfaite et un modèle qui croit parfois qu’un seul visage mérite toute une franchise.

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 prompts et les exemples sont adultes.

Ce modèle avait été annoncé il y a quelque temps sous le nom Chinese-Female-NSFW-LoRA-v1, avant de disparaître discrètement dans l’atelier. Il n’avait pas été oublié : il continuait simplement à découvrir de nouvelles façons d’échouer au contrôle qualité.

Cette version est le résultat de cette longue expérimentation. Elle conserve des limites visibles, et une deuxième version reposant sur des paramètres d’entraînement entièrement revus est déjà en cours d’apprentissage. La V1 est publiée comme un jalon utile mais imparfait, et non comme la preuve que le problème est résolu.

Les données d’entraînement proviennent d’une vaste collection de photographies d’adultes recueillies sur X. Elles comprennent des selfies, des clichés spontanés en intérieur et en extérieur, des scènes du quotidien ainsi que des interactions intimes entre partenaires adultes. Contrairement à Chinese Female Anatomy NSFW LoRA v8, cette expérience n’a pas été conçue en priorité pour maximiser la précision anatomique. Son objectif principal était d’apprendre l’irrégularité, l’immédiateté et la diversité de la photographie ordinaire.

Les exemples ci-dessous ont été générés avec les mêmes prompts que ceux utilisés pour Chinese Female Anatomy NSFW LoRA v8, ce qui permet une comparaison A/B directe entre les deux projets.

Couverture

Prompt

A quiet fine-art photograph portrays an adult nude Asian woman with subtle asymmetry and an introspective expression leaning lightly against a classical column with one knee relaxed beneath the arcades of a quiet market square before the shops open. A few damp strands of hair rest naturally against her face and shoulders. A waist-level side view follows the natural line of the gesture. Muted golden light reflects from old stone while the background remains cool. The treatment recalls a European art-photography monograph, quiet, tactile and observational. The gesture reads as a private moment observed with care and distance. Keep the gesture physically achievable and the relationship between body, gravity and support convincing.

Ce que ce LoRA apprend généralement

• Des cadrages spontanés, proches du selfie et de la photographie du quotidien
• Une plus grande variété de pièces ordinaires, d’objets, d’arrière-plans et de conditions d’éclairage
• Un rendu moins uniformément lissé qu’avec un jeu de données exclusivement réalisé en studio
• Des sujets adultes, avec des photographies en solo et des scènes intimes entre partenaires adultes dans les données d’entraînement
• Les irrégularités naturelles de la photographie — celles qui sont utiles, et parfois celles qui le sont moins

Détails de l’entraînement

Utilisation recommandée

• Pipeline recommandé : Qwen Image 2512 T2I
• Poids LoRA conseillé : 0.3–0.6 ; ne commencez pas au-dessus de 0.6
• Également testé : Qwen Image Edit 2511
• Déconseillé en premier choix : Qwen Image 2508 T2I, avec lequel le modèle manifeste plus facilement du surapprentissage et un effondrement de l’identité

Le LoRA a été entraîné sur la base 2508, mais il est généralement plus utile avec 2512 pour l’inférence. Ce flux de travail utilise deux bases différentes ; des artefacts de compatibilité peuvent donc apparaître ponctuellement. Commencez avec un poids faible et augmentez-le uniquement si le prompt nécessite une influence plus marquée du LoRA.

Limites connues

• Tendance au même visage : certains résultats convergent vers un visage familier, surtout avec un poids LoRA élevé ou des prompts courts et peu directifs.
• Biais lié aux filtres de beauté : une partie des photographies sources présentait un traitement esthétique visible. Cela peut produire une peau excessivement lisse, des visages homogénéisés ou un aspect cosmétique artificiel.
• Contamination par des filigranes : certaines images sources contenaient des filigranes visibles. Le modèle peut donc parfois reproduire du texte, des logos ou des artefacts graphiques qui y ressemblent.
• Erreurs anatomiques : l’anatomie n’était pas la priorité de ce jeu de données. Les mains, les membres, les articulations, l’anatomie sexuelle et les poses difficiles peuvent encore échouer.
• Surapprentissage avec 2508 : T2I 2508 peut reproduire la distribution d’entraînement de manière trop agressive. Qwen Image 2512 avec un poids plus faible constitue un point de départ plus sûr.
• Edit 2511 corrige, mais ne fait pas de magie : il peut aider à retoucher un visage ou une zone locale, sans pour autant effacer la mémorisation ni garantir la justesse anatomique.
• Entraînement long : après 100 000 étapes, le modèle a appris certaines choses avec une assurance admirable — y compris quelques-unes qu’il aurait peut-être dû examiner d’un œil plus critique.

Exemples

Avant d’entrer dans la collection d’intimité quotidienne du modèle et dans son interprétation parfois aventureuse de l’anatomie, commençons par quelques images qui se souviennent encore de la composition, de l’atmosphère et de cette idée un peu ancienne selon laquelle la photographie devrait conserver un peu de poésie.

Les exemples montrant des interactions intimes entre partenaires adultes ne sont volontairement pas inclus ici. L’intimité réelle reste une fonction que la réalité exécute mieux : aucun sampler, aucune échelle CFG et, les bons jours, nettement moins de membres dupliqués.

Exemple 1

Prompt

An adult Asian woman posing fully nude in a spacious old industrial loft, kneeling on both knees in a deep camel-pose backbend, torso arched dramatically backward, chest lifted, head tilted fully back, eyes closed, hips pushed forward, arms extended downward with her hands resting near her ankles, side-rear three-quarter view, full body visible. Her dark hair is tied in a low bun. Large black-framed grid windows fill the background, soft overcast daylight entering the room, blurred green foliage outside, pale weathered walls, wet reflective stone floor, muted neutral colors, shallow depth of field, soft natural skin texture, realistic anatomy, elegant fine-art nude photography, vertical composition, cinematic realism.

---

Exemple 2

Cette pose pourrait intéresser tout particulièrement la Fédération internationale de gymnastique — surtout parce que la gravité semble avoir présenté sa démission. Vous pouvez essayer le prompt ; tenter vous-même la pose est nettement moins recommandé.

Prompt

An adult Asian woman posing fully nude while performing an advanced contortion handstand in an empty industrial studio, balancing firmly on both hands with straight arms, shoulders stacked above her palms, torso curved into an extreme backbend, hips high above her head, one leg extended horizontally backward and the other bent deeply with the foot drawn toward her torso, side profile, full body visible. Her long dark hair hangs freely toward the floor as she looks downward with a focused expression. Minimal weathered white walls, two tall black-framed industrial windows on opposite sides, bare gray concrete floor, soft diffused daylight, subtle window light across the wall, muted neutral palette, realistic anatomy and muscle tension, natural skin texture, elegant athletic fine-art nude photography, shallow depth of field, vertical composition, cinematic realism.

---

Exemple 3

Prompt

A documentary-inspired art photograph records an adult nude Asian woman with realistic proportions and natural skin texture holding a wide kneeling backbend with the chest lifted, hands reaching toward the heels and pelvis presented obliquely within a derelict chapel where pale daylight enters through clear broken panes. A slightly elevated diagonal view shows the pose without flattening the pelvic structure, presenting the pelvic anatomy clearly while preserving a believable spine, hips and leg alignment. Diffused skylight and reflected water produce quiet moving highlights. The aesthetic is minimal and introspective, favoring weather, material texture and human form over glamour. The scene feels candid, meditative and grounded in ordinary physical reality. Preserve full-body continuity, natural hands and feet, believable weight distribution and anatomically consistent hips.

---

<details> <summary><strong>🔞 Cliquer pour afficher les exemples anatomiques explicites 4 à 7

La section suivante contient des études anatomiques adultes explicites. Elle est repliée par défaut afin que son affichage résulte d’un choix délibéré plutôt que d’un changement de décor inattendu. </strong></summary>

---

Exemple 4 — Les feuilles d’un même arbre

Un prompt est entré ; quatre interprétations en sont sorties. Comme des feuilles photographiées sous des vents différents, les résultats partagent un air de famille sans s’accorder sur chaque ligne. Le sampler considère apparemment l’uniformité anatomique comme une suggestion plutôt que comme un contrat.

Prompt commun

Close-up of an adult Asian woman's vulva with visible labia majora, labia minora, and clitoral hood. The vulva is surrounded by dark pubic hair. Two hands with light blue and white manicured nails gently part the labia majora. The woman is wearing beige lace underwear, partially visible at the edges of the frame. The background is soft, out-of-focus beige fabric. Soft even lighting, realistic high-resolution photograph, shallow depth of field.

---

Exemple 5

Prompt commun

An Asian woman in her thirties, fully nude, kneeling with her hands on the ground, legs spread apart and hips raised, looking back at the camera, direct-flash photography, rear view.

---

Exemple 6

Prompt

an adult Asian woman is nude and positioned on all fours on a bed with a yellow floral-patterned sheet. Her back is arched and her buttocks are prominently visible. Two pillows are on the bed: one pink with a cartoon character and one red and purple with floral patterns. The room has white walls and a light switch is visible on the right wall. Even natural lighting and realistic photographic style.

Exemple 7

Prompt

an adult Asian woman is lying in a bathtub filled with soapy water. She is nude, with her legs spread and knees bent. She has long dark hair, some of which is wet and draped over her shoulders. Her right arm is bent and her hand is near her hair. The bathtub is white and the surrounding walls are tiled with light-colored tiles. Bright even lighting and clear realistic photographic style.

</details>

---

Exemple 8

Prompt

Desaturated cinematic figure study of an adult Asian woman, fully nude, kneeling with torso lowered and head turned back, outdoors in a deserted gravel pit beneath raw excavated slopes, during a misty daytime with weak diffuse illumination, the damaged terrain lending the composition a restrained desolation, desaturated mineral colors and low-contrast distant detail, photographed in a centered vertical gallery composition, high-resolution body detail and neutral color science, the environment deliberately less prominent, no jewelry, no clothing.

---

Exemple 9

Prompt

This intimate documentary-style image records an adult nude Asian woman integrated quietly into the surrounding environment, sitting on the floor beside an old armchair with knees drawn loosely toward her body inside a spare artist's loft with limewashed walls and tall factory windows. A few damp strands of hair rest naturally against her face and shoulders. The camera remains near floor level but avoids theatrical distortion. A pale morning glow enters from an adjoining room and creates long tonal transitions. The image is rendered in luminous black and white with a long grayscale and delicate silver-grain texture. The emotional tone is reflective and slightly melancholic. Preserve full-body continuity, natural hands and feet, plausible weight distribution and understated expression.

---

Une dernière image avant de partir

Aucun prompt n’accompagne cette image. La lumière matinale du Montana n’arrive ni avec une zone de texte, ni avec un numéro de seed, ni avec un bouton de régénération bien pratique.

---

🌍 Une ultime recommandation hautement technique

Une recommandation sérieuse : ne passez pas toute la journée à négocier avec un sampler. Fermez le notebook, partez marcher longtemps, rencontrez de vraies personnes et laissez la lumière imparfaite du jour remettre vos yeux à zéro. Le monde offre davantage de visages, de gestes, d’histoires et d’accidents de lumière que n’importe quel jeu d’entraînement — le plus souvent sans consommer 96 Go de VRAM.

---

🎞️ Collection nettoyée de prompts issus de X

Ce dépôt comprend twitter_x_cleaned_prompts.txt, qui contient 9 310 prompts nettoyés issus des légendes d’entraînement appariées. Les fragments invalides ont été supprimés, la formulation a été normalisée afin d’indiquer explicitement que les sujets sont adultes, et les images sources originales ne sont pas incluses.

Certains prompts restent imparfaits, car la réalité est désordonnée et la génération automatique de légendes n’a encore jamais rencontré un objet qu’elle ne puisse décrire avec une assurance superflue.

---

🚀 Code de référence — Qwen-Image-2512 T2I

Le fichier téléchargeable reference.py charge Qwen-Image-2512 en BF16, fusionne l’adaptateur Lightning officiel de la communauté et applique ce LoRA. Un GPU CUDA disposant d’une grande quantité de mémoire est recommandé.

Installation

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary><strong>▶ Afficher le code Python de référence standard</strong></summary>

reference.py

"""Qwen Image 2512 + Chinese Female NSFW X 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/Chinese-Female-NSFW-X-LoRA-v1"
LORA_WEIGHT_NAME = "twitter.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.5


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="twitter_x_v1",
)
pipe.set_adapters(["twitter_x_v1"], adapter_weights=[LORA_SCALE])

prompt = (
    "zl_cn_female_x_v1, an adult Asian woman in a candid environmental "
    "photograph, natural expression, realistic skin texture, documentary light"
)
negative_prompt = (
    "child, minor, underage, low quality, blurry, distorted anatomy, malformed "
    "hands, extra limbs, duplicated body parts, 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_nsfw_x_v1.png")

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

reference_low_vram.py

"""Low-VRAM Qwen Image 2512 + Chinese Female NSFW X 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/Chinese-Female-NSFW-X-LoRA-v1"
LORA_WEIGHT_NAME = "twitter.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.5


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="twitter_x_v1",
)
pipe.set_adapters(["twitter_x_v1"], adapter_weights=[LORA_SCALE])

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

prompt = (
    "zl_cn_female_x_v1, an adult Asian woman in a candid environmental "
    "photograph, natural expression, realistic skin texture, documentary light"
)
negative_prompt = (
    "child, minor, underage, low quality, blurry, distorted anatomy, malformed "
    "hands, extra limbs, duplicated body parts, 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_nsfw_x_v1_low_vram.png")

</details>

---

🔞 Utilisation responsable

Ce modèle est destiné exclusivement à des créations légales, consensuelles et réservées aux adultes. Ne l’utilisez pas pour :

• créer ou sexualiser des mineurs ou des personnes paraissant mineures
• 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 appartient aux utilisateurs de s’assurer que leurs prompts, leurs sources et leurs contenus générés respectent les lois applicables et les règles des plateformes.

---

Licence et droits d’auteur

Ce LoRA peut uniquement être utilisé 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, veuillez contacter l’auteur.

---

Aperçu de la V2

La V2 est déjà en cours d’entraînement avec une configuration entièrement revue. Son objectif est d’améliorer le réalisme anatomique sans sacrifier une composition crédible ni un éclairage naturel, tout en corrigeant la généralisation limitée de la V1. Autrement dit : moins de visages répétés, moins de raccourcis mémorisés et, espérons-le, moins de moments où l’anatomie tente une refonte non autorisée d’elle-même.

Voici deux premiers aperçus des résultats de la V2 :

Aperçu V2 1

Aperçu V2 2

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

---

LoRA NSFW X de mujeres chinas v1

Fotografías cotidianas, realidad imperfecta y un modelo que a veces cree que un solo rostro merece una franquicia completa.

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

Este modelo se presentó hace tiempo como Chinese-Female-NSFW-LoRA-v1 y luego desapareció discretamente en el taller. No cayó en el olvido; simplemente siguió descubriendo nuevas maneras de reprobar el control de calidad.

Esta versión es el resultado de ese largo experimento. Todavía conserva limitaciones visibles, y ya se está entrenando una segunda versión con parámetros completamente revisados. La V1 se publica como un punto de control útil pero imperfecto, no como una declaración de que el problema ya está resuelto.

Los datos de entrenamiento provienen de una amplia colección de fotografías de adultos recopiladas en X. Incluyen selfies, imágenes casuales en interiores y exteriores, situaciones cotidianas e interacciones íntimas entre parejas adultas. A diferencia de Chinese Female Anatomy NSFW LoRA v8, este experimento no fue diseñado principalmente para maximizar la precisión anatómica. Su objetivo central era aprender la irregularidad, la inmediatez y la variedad de la fotografía cotidiana.

Los ejemplos siguientes se generaron con los mismos prompts utilizados en Chinese Female Anatomy NSFW LoRA v8, lo que permite comparar ambos proyectos directamente mediante una prueba A/B.

Portada

Prompt

A quiet fine-art photograph portrays an adult nude Asian woman with subtle asymmetry and an introspective expression leaning lightly against a classical column with one knee relaxed beneath the arcades of a quiet market square before the shops open. A few damp strands of hair rest naturally against her face and shoulders. A waist-level side view follows the natural line of the gesture. Muted golden light reflects from old stone while the background remains cool. The treatment recalls a European art-photography monograph, quiet, tactile and observational. The gesture reads as a private moment observed with care and distance. Keep the gesture physically achievable and the relationship between body, gravity and support convincing.

Lo que este LoRA suele aprender

• Encuadres espontáneos, similares a selfies y propios de la fotografía cotidiana
• Una mayor variedad de habitaciones comunes, objetos, fondos y condiciones de iluminación
• Una presentación menos uniformemente pulida que la de un conjunto de datos exclusivamente de estudio
• Sujetos adultos, incluidas fotografías individuales y escenas íntimas entre parejas adultas dentro de los datos de entrenamiento
• La inconsistencia natural de la fotografía: la clase útil y, en ocasiones, la que no ayuda tanto

Detalles del entrenamiento

Uso recomendado

• Pipeline recomendado: Qwen Image 2512 T2I
• Peso LoRA sugerido: 0.3–0.6; no empieces por encima de 0.6
• También probado: Qwen Image Edit 2511
• No se recomienda como primera opción: Qwen Image 2508 T2I, donde el modelo muestra con mayor facilidad sobreajuste y colapso de identidad

El LoRA se entrenó sobre la base 2508, pero en general resulta más útil con 2512 durante la inferencia. Este flujo de trabajo cruza dos bases, por lo que pueden aparecer artefactos ocasionales de compatibilidad. Empieza con un peso bajo y auméntalo solo cuando el prompt necesite una mayor influencia del LoRA.

Limitaciones conocidas

• Tendencia al mismo rostro: algunos resultados convergen hacia un rostro familiar, especialmente con pesos LoRA altos o prompts breves y poco definidos.
• Sesgo de filtros de belleza: parte de las fotografías originales tenía retoques de belleza visibles. Esto puede producir piel demasiado lisa, homogeneización facial o una apariencia cosmética artificial.
• Contaminación por marcas de agua: algunas imágenes originales contenían marcas de agua visibles. Por ello, el modelo puede reproducir ocasionalmente texto, logotipos o artefactos gráficos similares.
• Errores anatómicos: este conjunto de datos no priorizó la anatomía. Las manos, extremidades, articulaciones, anatomía sexual y poses difíciles aún pueden fallar.
• Sobreajuste en 2508: T2I 2508 puede reproducir la distribución de entrenamiento con demasiada intensidad. Qwen Image 2512 con un peso menor es el punto de partida más seguro.
• Edit 2511 corrige, pero no hace magia: puede ayudar a revisar un rostro o una región concreta, pero no elimina la memorización ni garantiza la corrección anatómica.
• Entrenamiento prolongado: tras 100.000 pasos, el modelo aprendió algunas cosas con una confianza admirable, incluidas varias que quizá debió cuestionar con un poco más de cuidado.

Ejemplos

Antes de entrar en la colección de intimidad cotidiana del modelo y en su interpretación ocasionalmente aventurera de la anatomía, comencemos con algunas imágenes que todavía recuerdan la composición, la atmósfera y esa idea algo anticuada de que la fotografía debería conservar un poco de poesía.

Los ejemplos con interacción íntima entre parejas adultas se omiten deliberadamente. La intimidad real sigue siendo una función que la realidad ejecuta mejor: sin sampler, sin escala CFG y, en un buen día, con bastantes menos extremidades duplicadas.

Ejemplo 1

Prompt

An adult Asian woman posing fully nude in a spacious old industrial loft, kneeling on both knees in a deep camel-pose backbend, torso arched dramatically backward, chest lifted, head tilted fully back, eyes closed, hips pushed forward, arms extended downward with her hands resting near her ankles, side-rear three-quarter view, full body visible. Her dark hair is tied in a low bun. Large black-framed grid windows fill the background, soft overcast daylight entering the room, blurred green foliage outside, pale weathered walls, wet reflective stone floor, muted neutral colors, shallow depth of field, soft natural skin texture, realistic anatomy, elegant fine-art nude photography, vertical composition, cinematic realism.

---

Ejemplo 2

Esta pose podría interesarle especialmente a la Federación Internacional de Gimnasia, principalmente porque la gravedad parece haber presentado su renuncia. Puedes probar el prompt; intentar la pose en persona es mucho menos recomendable.

Prompt

An adult Asian woman posing fully nude while performing an advanced contortion handstand in an empty industrial studio, balancing firmly on both hands with straight arms, shoulders stacked above her palms, torso curved into an extreme backbend, hips high above her head, one leg extended horizontally backward and the other bent deeply with the foot drawn toward her torso, side profile, full body visible. Her long dark hair hangs freely toward the floor as she looks downward with a focused expression. Minimal weathered white walls, two tall black-framed industrial windows on opposite sides, bare gray concrete floor, soft diffused daylight, subtle window light across the wall, muted neutral palette, realistic anatomy and muscle tension, natural skin texture, elegant athletic fine-art nude photography, shallow depth of field, vertical composition, cinematic realism.

---

Ejemplo 3

Prompt

A documentary-inspired art photograph records an adult nude Asian woman with realistic proportions and natural skin texture holding a wide kneeling backbend with the chest lifted, hands reaching toward the heels and pelvis presented obliquely within a derelict chapel where pale daylight enters through clear broken panes. A slightly elevated diagonal view shows the pose without flattening the pelvic structure, presenting the pelvic anatomy clearly while preserving a believable spine, hips and leg alignment. Diffused skylight and reflected water produce quiet moving highlights. The aesthetic is minimal and introspective, favoring weather, material texture and human form over glamour. The scene feels candid, meditative and grounded in ordinary physical reality. Preserve full-body continuity, natural hands and feet, believable weight distribution and anatomically consistent hips.

---

<details> <summary><strong>🔞 Haz clic para mostrar los ejemplos anatómicos explícitos 4–7

La siguiente sección contiene estudios anatómicos explícitos de adultos. Está plegada de forma predeterminada para que verla sea una decisión consciente y no un cambio de escenario inesperado. </strong></summary>

---

Ejemplo 4 — Hojas del mismo árbol

Entró un prompt y salieron cuatro interpretaciones. Como hojas fotografiadas con vientos distintos, los resultados comparten un parecido familiar sin coincidir en cada línea. Al parecer, el sampler considera la uniformidad anatómica una sugerencia y no un contrato.

Prompt compartido

Close-up of an adult Asian woman's vulva with visible labia majora, labia minora, and clitoral hood. The vulva is surrounded by dark pubic hair. Two hands with light blue and white manicured nails gently part the labia majora. The woman is wearing beige lace underwear, partially visible at the edges of the frame. The background is soft, out-of-focus beige fabric. Soft even lighting, realistic high-resolution photograph, shallow depth of field.

---

Ejemplo 5

Prompt compartido

An Asian woman in her thirties, fully nude, kneeling with her hands on the ground, legs spread apart and hips raised, looking back at the camera, direct-flash photography, rear view.

---

Ejemplo 6

Prompt

an adult Asian woman is nude and positioned on all fours on a bed with a yellow floral-patterned sheet. Her back is arched and her buttocks are prominently visible. Two pillows are on the bed: one pink with a cartoon character and one red and purple with floral patterns. The room has white walls and a light switch is visible on the right wall. Even natural lighting and realistic photographic style.

Ejemplo 7

Prompt

an adult Asian woman is lying in a bathtub filled with soapy water. She is nude, with her legs spread and knees bent. She has long dark hair, some of which is wet and draped over her shoulders. Her right arm is bent and her hand is near her hair. The bathtub is white and the surrounding walls are tiled with light-colored tiles. Bright even lighting and clear realistic photographic style.

</details>

---

Ejemplo 8

Prompt

Desaturated cinematic figure study of an adult Asian woman, fully nude, kneeling with torso lowered and head turned back, outdoors in a deserted gravel pit beneath raw excavated slopes, during a misty daytime with weak diffuse illumination, the damaged terrain lending the composition a restrained desolation, desaturated mineral colors and low-contrast distant detail, photographed in a centered vertical gallery composition, high-resolution body detail and neutral color science, the environment deliberately less prominent, no jewelry, no clothing.

---

Ejemplo 9

Prompt

This intimate documentary-style image records an adult nude Asian woman integrated quietly into the surrounding environment, sitting on the floor beside an old armchair with knees drawn loosely toward her body inside a spare artist's loft with limewashed walls and tall factory windows. A few damp strands of hair rest naturally against her face and shoulders. The camera remains near floor level but avoids theatrical distortion. A pale morning glow enters from an adjoining room and creates long tonal transitions. The image is rendered in luminous black and white with a long grayscale and delicate silver-grain texture. The emotional tone is reflective and slightly melancholic. Preserve full-body continuity, natural hands and feet, plausible weight distribution and understated expression.

---

Una última imagen antes de irte

Esta imagen no incluye prompt. La luz matinal de Montana no llega con una caja de texto, un número de seed ni un cómodo botón para regenerar.

---

🌍 Una última recomendación altamente técnica

Una recomendación seria: no pases todo el día negociando con un sampler. Cierra el notebook, sal a caminar, conoce gente real y deja que la luz imperfecta del día reajuste tu mirada. El mundo ofrece más rostros, gestos, historias y accidentes de luz que cualquier conjunto de entrenamiento, normalmente sin consumir 96 GB de VRAM.

---

🎞️ Colección depurada de prompts de X

Este repositorio incluye twitter_x_cleaned_prompts.txt, con 9.310 prompts depurados derivados de las leyendas de entrenamiento emparejadas. Se eliminaron los fragmentos inválidos, se normalizó la redacción para dejar explícito que los sujetos son adultos y no se incluyen las imágenes originales.

Algunos prompts siguen siendo imperfectos porque la realidad es desordenada y el etiquetado automático nunca ha conocido un objeto que no pudiera describir con una confianza innecesaria.

---

🚀 Código de referencia — Qwen-Image-2512 T2I

El archivo descargable reference.py carga Qwen-Image-2512 en BF16, fusiona el adaptador Lightning oficial de la comunidad y aplica este LoRA. Se recomienda una GPU CUDA con gran capacidad de memoria.

Instalación

pip install -U diffusers transformers accelerate peft safetensors

<details> <summary><strong>▶ Mostrar el código Python de referencia estándar</strong></summary>

reference.py

"""Qwen Image 2512 + Chinese Female NSFW X 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/Chinese-Female-NSFW-X-LoRA-v1"
LORA_WEIGHT_NAME = "twitter.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.5


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="twitter_x_v1",
)
pipe.set_adapters(["twitter_x_v1"], adapter_weights=[LORA_SCALE])

prompt = (
    "zl_cn_female_x_v1, an adult Asian woman in a candid environmental "
    "photograph, natural expression, realistic skin texture, documentary light"
)
negative_prompt = (
    "child, minor, underage, low quality, blurry, distorted anatomy, malformed "
    "hands, extra limbs, duplicated body parts, 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_nsfw_x_v1.png")

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

reference_low_vram.py

"""Low-VRAM Qwen Image 2512 + Chinese Female NSFW X 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/Chinese-Female-NSFW-X-LoRA-v1"
LORA_WEIGHT_NAME = "twitter.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.5


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="twitter_x_v1",
)
pipe.set_adapters(["twitter_x_v1"], adapter_weights=[LORA_SCALE])

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

prompt = (
    "zl_cn_female_x_v1, an adult Asian woman in a candid environmental "
    "photograph, natural expression, realistic skin texture, documentary light"
)
negative_prompt = (
    "child, minor, underage, low quality, blurry, distorted anatomy, malformed "
    "hands, extra limbs, duplicated body parts, 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_nsfw_x_v1_low_vram.png")

</details>

---

🔞 Uso responsable

Este modelo está destinado exclusivamente a trabajos creativos legales, consensuados y para adultos. No lo utilices para:

• crear o sexualizar a menores o personas de apariencia juvenil
• crear contenido sexual no consensuado o de explotación
• representar a una persona real identificable sin permiso
• acosar, amenazar, explotar o difamar a otra persona
• crear contenido ilegal o evadir las reglas de una plataforma

Cada usuario es responsable 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. Consulta el archivo LICENSE del repositorio para conocer todos los términos.

Para consultas sobre licencias comerciales, ponte en contacto con el autor.

---

Avance de la V2

La V2 ya se está entrenando con una configuración completamente revisada. Su objetivo es mejorar el realismo anatómico sin sacrificar una composición creíble ni la iluminación natural, al tiempo que corrige la generalización limitada de la V1. En otras palabras: menos rostros repetidos, menos atajos memorizados y, con suerte, menos momentos en los que la anatomía intente rediseñarse sin autorización.

Estos son dos primeros resultados preliminares de la V2:

Avance V2 1

Avance V2 2

</details>

---

Chinese Female NSFW X LoRA v1

Everyday photographs, imperfect reality, and a model that occasionally believes one face deserves an entire franchise.

Adult content warning: This repository contains explicit AI-generated imagery and is intended only for adults. All people described in the prompts and examples are adults.

This model was previewed some time ago as Chinese-Female-NSFW-LoRA-v1 and then quietly disappeared into the workshop. It was not forgotten; it simply kept finding new ways to fail quality control.

This release is the result of that long experiment. It still has visible limitations, and a second version with completely revised training parameters is already in training. V1 is published as a useful, imperfect checkpoint rather than a declaration that the problem has been solved.

The training data came from a large collection of adult photographs gathered from X. It includes selfies, casual indoor and outdoor photographs, everyday situations, and intimate interaction between adult partners. Unlike Chinese Female Anatomy NSFW LoRA v8, this experiment was not primarily designed to maximize anatomical accuracy. Its main goal was to learn the irregularity, immediacy, and variety of ordinary photography.

The examples below were generated with the same prompts used for Chinese Female Anatomy NSFW LoRA v8, making the two projects useful for direct A/B comparison.

Cover

Prompt

A quiet fine-art photograph portrays an adult nude Asian woman with subtle asymmetry and an introspective expression leaning lightly against a classical column with one knee relaxed beneath the arcades of a quiet market square before the shops open. A few damp strands of hair rest naturally against her face and shoulders. A waist-level side view follows the natural line of the gesture. Muted golden light reflects from old stone while the background remains cool. The treatment recalls a European art-photography monograph, quiet, tactile and observational. The gesture reads as a private moment observed with care and distance. Keep the gesture physically achievable and the relationship between body, gravity and support convincing.

What this LoRA tends to learn

• Candid, selfie-like, and everyday photographic framing
• A broader range of ordinary rooms, objects, backgrounds, and lighting conditions
• Less uniformly polished presentation than a studio-only dataset
• Adult subjects, including both solo photographs and intimate adult-partner scenes in the training data
• Natural photographic inconsistency—the useful kind, and occasionally the less useful kind

Training details

Recommended use

• Recommended pipeline: Qwen Image 2512 T2I
• Suggested LoRA weight: 0.3–0.6; do not begin above 0.6
• Also tested: Qwen Image Edit 2511
• Not recommended as the first choice: Qwen Image 2508 T2I, where the model more readily shows overfitting and identity collapse

The LoRA was trained on the 2508 base but is generally more useful on 2512 for inference. This is a cross-base workflow, so occasional compatibility artifacts are possible. Start with a low weight and increase it only when the prompt needs more influence from the LoRA.

Known limitations

• Same-face tendency: Some outputs converge toward a familiar face, especially at higher LoRA weights or with short, weak prompts.
• Beauty-filter bias: A portion of the source photographs had visible beauty processing. This can produce overly smooth skin, facial homogenization, or an artificial cosmetic look.
• Watermark contamination: Some source images contained visible watermarks. The model may therefore occasionally reproduce watermark-like text, logos, or graphic artifacts.
• Anatomical errors: This was not an anatomy-first dataset. Hands, limbs, joints, sexual anatomy, and difficult poses can still fail.
• Overfitting on 2508: T2I 2508 can reproduce the training distribution too aggressively. Qwen Image 2512 with a lower weight is the safer starting point.
• Edit 2511 is corrective, not magical: It can help revise a face or local region, but it does not erase memorization or guarantee anatomical correctness.
• Long training: At 100,000 steps, the model has learned some things with admirable confidence—including a few things it should perhaps have questioned more carefully.

Examples

Before we enter the model's collection of everyday intimacy and its occasionally adventurous interpretation of anatomy, let us begin with a few images that still remember composition, atmosphere, and the old-fashioned idea that photography should have a little poetry.

Examples involving intimate interaction between adult partners are deliberately not included here. Real intimacy is still one feature reality implements better—no sampler, no CFG scale, and, on a good day, considerably fewer duplicated limbs.

Example 1

Prompt

An adult Asian woman posing fully nude in a spacious old industrial loft, kneeling on both knees in a deep camel-pose backbend, torso arched dramatically backward, chest lifted, head tilted fully back, eyes closed, hips pushed forward, arms extended downward with her hands resting near her ankles, side-rear three-quarter view, full body visible. Her dark hair is tied in a low bun. Large black-framed grid windows fill the background, soft overcast daylight entering the room, blurred green foliage outside, pale weathered walls, wet reflective stone floor, muted neutral colors, shallow depth of field, soft natural skin texture, realistic anatomy, elegant fine-art nude photography, vertical composition, cinematic realism.

---

Example 2

This pose may be of particular interest to the International Gymnastics Federation—mostly because gravity appears to have submitted its resignation. You are welcome to try the prompt. Trying the pose yourself is considerably less recommended.

Prompt

An adult Asian woman posing fully nude while performing an advanced contortion handstand in an empty industrial studio, balancing firmly on both hands with straight arms, shoulders stacked above her palms, torso curved into an extreme backbend, hips high above her head, one leg extended horizontally backward and the other bent deeply with the foot drawn toward her torso, side profile, full body visible. Her long dark hair hangs freely toward the floor as she looks downward with a focused expression. Minimal weathered white walls, two tall black-framed industrial windows on opposite sides, bare gray concrete floor, soft diffused daylight, subtle window light across the wall, muted neutral palette, realistic anatomy and muscle tension, natural skin texture, elegant athletic fine-art nude photography, shallow depth of field, vertical composition, cinematic realism.

---

Example 3

Prompt

A documentary-inspired art photograph records an adult nude Asian woman with realistic proportions and natural skin texture holding a wide kneeling backbend with the chest lifted, hands reaching toward the heels and pelvis presented obliquely within a derelict chapel where pale daylight enters through clear broken panes. A slightly elevated diagonal view shows the pose without flattening the pelvic structure, presenting the pelvic anatomy clearly while preserving a believable spine, hips and leg alignment. Diffused skylight and reflected water produce quiet moving highlights. The aesthetic is minimal and introspective, favoring weather, material texture and human form over glamour. The scene feels candid, meditative and grounded in ordinary physical reality. Preserve full-body continuity, natural hands and feet, believable weight distribution and anatomically consistent hips.

---

<details> <summary><strong>🔞 Click to reveal explicit anatomy examples 4–7

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. </strong></summary>

---

Example 4 — Leaves From the Same Tree

One prompt went in; four interpretations came out. Like leaves photographed in different wind, the results share a family resemblance without agreeing on every line. The sampler, apparently, regards anatomical uniformity as a suggestion rather than a contract.

Shared Prompt

Close-up of an adult Asian woman's vulva with visible labia majora, labia minora, and clitoral hood. The vulva is surrounded by dark pubic hair. Two hands with light blue and white manicured nails gently part the labia majora. The woman is wearing beige lace underwear, partially visible at the edges of the frame. The background is soft, out-of-focus beige fabric. Soft even lighting, realistic high-resolution photograph, shallow depth of field.

---

Example 5

Shared Prompt

An Asian woman in her thirties, fully nude, kneeling with her hands on the ground, legs spread apart and hips raised, looking back at the camera, direct-flash photography, rear view.

---

Example 6

Prompt

an adult Asian woman is nude and positioned on all fours on a bed with a yellow floral-patterned sheet. Her back is arched and her buttocks are prominently visible. Two pillows are on the bed: one pink with a cartoon character and one red and purple with floral patterns. The room has white walls and a light switch is visible on the right wall. Even natural lighting and realistic photographic style.

Example 7

Prompt

an adult Asian woman is lying in a bathtub filled with soapy water. She is nude, with her legs spread and knees bent. She has long dark hair, some of which is wet and draped over her shoulders. Her right arm is bent and her hand is near her hair. The bathtub is white and the surrounding walls are tiled with light-colored tiles. Bright even lighting and clear realistic photographic style.

</details>

---

Example 8

Prompt

Desaturated cinematic figure study of an adult Asian woman, fully nude, kneeling with torso lowered and head turned back, outdoors in a deserted gravel pit beneath raw excavated slopes, during a misty daytime with weak diffuse illumination, the damaged terrain lending the composition a restrained desolation, desaturated mineral colors and low-contrast distant detail, photographed in a centered vertical gallery composition, high-resolution body detail and neutral color science, the environment deliberately less prominent, no jewelry, no clothing.

---

Example 9

Prompt

This intimate documentary-style image records an adult nude Asian woman integrated quietly into the surrounding environment, sitting on the floor beside an old armchair with knees drawn loosely toward her body inside a spare artist's loft with limewashed walls and tall factory windows. A few damp strands of hair rest naturally against her face and shoulders. The camera remains near floor level but avoids theatrical distortion. A pale morning glow enters from an adjoining room and creates long tonal transitions. The image is rendered in luminous black and white with a long grayscale and delicate silver-grain texture. The emotional tone is reflective and slightly melancholic. Preserve full-body continuity, natural hands and feet, plausible weight distribution and understated expression.

---

One More View Before You Go

No prompt accompanies this image. Morning light in Montana does not arrive with a text box, a seed number, or a convenient regenerate button.

---

🌍 One Last Highly Technical Recommendation

One serious recommendation: do not spend the whole day negotiating with a sampler. Close the notebook, take a long walk, meet real people, and let imperfect daylight reset your eyes. The world offers more faces, gestures, stories, and accidents of light than any training set—usually without consuming 96 GB of VRAM.

---

🎞️ Cleaned X Prompt Collection

This repository includes twitter_x_cleaned_prompts.txt, containing 9,310 cleaned prompts derived from the paired training captions. Invalid fragments were removed, wording was normalized to make adult status explicit, and the original source images are not included.

Some prompts remain imperfect because reality is untidy and automatic captioning has never met an object it could not describe with unnecessary confidence.

---

🚀 Reference Code — Qwen-Image-2512 T2I

The downloadable reference.py loads Qwen-Image-2512 in BF16, fuses the official community Lightning adapter, and applies this LoRA. A high-memory CUDA GPU is recommended.

Installation

pip install -U diffusers transformers accelerate peft safetensors

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

reference.py

"""Qwen Image 2512 + Chinese Female NSFW X 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/Chinese-Female-NSFW-X-LoRA-v1"
LORA_WEIGHT_NAME = "twitter.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.5


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="twitter_x_v1",
)
pipe.set_adapters(["twitter_x_v1"], adapter_weights=[LORA_SCALE])

prompt = (
    "zl_cn_female_x_v1, an adult Asian woman in a candid environmental "
    "photograph, natural expression, realistic skin texture, documentary light"
)
negative_prompt = (
    "child, minor, underage, low quality, blurry, distorted anatomy, malformed "
    "hands, extra limbs, duplicated body parts, 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_nsfw_x_v1.png")

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

reference_low_vram.py

"""Low-VRAM Qwen Image 2512 + Chinese Female NSFW X 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/Chinese-Female-NSFW-X-LoRA-v1"
LORA_WEIGHT_NAME = "twitter.safetensors"

DTYPE = torch.bfloat16
LORA_SCALE = 0.5


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="twitter_x_v1",
)
pipe.set_adapters(["twitter_x_v1"], adapter_weights=[LORA_SCALE])

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

prompt = (
    "zl_cn_female_x_v1, an adult Asian woman in a candid environmental "
    "photograph, natural expression, realistic skin texture, documentary light"
)
negative_prompt = (
    "child, minor, underage, low quality, blurry, distorted anatomy, malformed "
    "hands, extra limbs, duplicated body parts, 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_nsfw_x_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, please contact the author.

---

V2 Preview

V2 is already in training with a completely revised setup. Its goal is to improve anatomical realism without sacrificing believable composition and natural lighting, while addressing V1's limited generalization. In other words: fewer repeated faces, fewer memorized shortcuts, and hopefully fewer moments when anatomy attempts an unauthorized redesign of itself.

These are two early V2 preview results:

V2 Preview 1

V2 Preview 2