Downloads · 30 days
3.6K
31% of all-time downloads
perpetual3x/Qwen-Image-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2
Qwen-Image-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2 is a text-to-image model from perpetual3x. Use it when you need an image from a text prompt. It is set up for diffusers. The card lists the license as other.
<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:/…
Downloads · 30 days
3.6K
31% of all-time downloads
All-time downloads
11.4K
Public
Repo size
610 MB
Likes
11
Public
Click a slice to open those files.
.safetensors590 MB · 97%
From the Hugging Face model README
🇬🇧 English — displayed below

Avertissement — contenu pour adultes : ce dépôt contient des images explicites générées par IA et s’adresse exclusivement aux adultes. Toutes les personnes décrites dans les exemples et les prompts de référence sont adultes.
De nombreux LoRA texte-vers-image savent restituer les reflets chromés d’une moto, chaque grain de sable et la météo émotionnelle précise d’un orage imminent — puis arrivent au bassin et déposent un espace réservé. Ce LoRA a été entraîné pour réparer cette lacune particulière dans des générations par ailleurs impressionnantes.
Entraîner des LoRA ressemble étonnamment à la course de fond. J’entraîne des modèles, mais je cours aussi environ 160 miles par mois — soit 5,3 miles, ou à peu près 8 500 à 11 500 pas humains par jour selon la foulée. Mes jambes peuvent donc enregistrer davantage de pas en une journée ordinaire que ce LoRA pendant ses 8 000 étapes d’entraînement. Pourtant, ni un coureur ni un modèle ne progressent par la seule accumulation des pas. La fatigue finit par révéler les défauts de forme d’un coureur ; un modèle peut cacher les siens derrière une graine flatteuse. Voilà pourquoi la validation compte davantage que le kilométrage.
Il s’agit d’un LoRA d’assistance anatomique, et non d’un modèle autonome complet de personnage, d’identité, de style ou de composition. Il fonctionne au mieux aux côtés d’un autre LoRA NSFW qui fournit déjà la personne, le langage visuel, la scène ou le caractère photographique. Considérez le LoRA principal comme l’architecte, et cet adapter comme le spécialiste qui demande discrètement si le plan de plomberie a bien été inclus.
Vous souvenez-vous du Pixiu(貔貅), la créature mythique qui attire la richesse et que j’évoquais dans la fiche du modèle Triggerless v2 ? La tradition affirme qu’il n’a pas d’anus — excellent pour retenir les trésors, moins utile comme référence anatomique. Et non, rester assis toute la journée ne crée pas d’anus supplémentaires. Cela peut toutefois favoriser le genre de problèmes anorectaux qui donnent au premier l’impression d’être en sureffectif — une raison de plus de se lever et de sortir. Nous sommes humains, mais toujours des animaux façonnés par le mouvement, la lumière du jour, la distance et le terrain. Gardez quelque chose d’élémentaire dans votre vie : sortez, marchez, courez, gravissez une colline, et laissez le corps se souvenir de ce que les écrans — et parfois les modèles — savent remarquablement bien oublier.
Le but de meilleurs outils n’est pas de nous garder plus longtemps devant eux. Entraînez le modèle, inspectez les détails, puis fermez l’écran et revenez à l’image plus vaste : un corps qui bouge, un paysage impossible à compresser dans un prompt, et une vie dont la valeur ne se mesure pas seulement en étapes.
La Triggerless v2 a été conçue pour agir discrètement à partir d’un langage descriptif ordinaire. C’est pratique lorsque la correction anatomique doit rester disponible en permanence, mais certains workflows ont besoin d’un interrupteur plus net : activer volontairement la correction, renforcer son influence et la tenir à l’écart lorsqu’elle n’est pas souhaitée.
C’est la raison d’être de la Triggered v2. Elle ne remplace pas la Triggerless v2 et ne s’appelle pas v3 simplement parce que le mot déclencheur n’a pas soudainement obtenu une promotion. Les deux modèles appartiennent à la même famille v2 validée : l’un attend poliment en arrière-plan, l’autre apporte la boîte à outils uniquement lorsqu’on l’appelle par son nom.
Le mot déclencheur requis est :
zl_female_anatomy_fix
| Dimension | Version sans déclencheur | Version avec déclencheur |
|---|---|---|
| Aspect naturel | Plus organique ; la correction paraît naturellement intégrée | Plus affirmée ; la zone corrigée peut parfois sembler volontairement mise en avant |
| Visibilité anatomique | Les angles difficiles peuvent encore omettre l’anus | Plus susceptible de produire l’anatomie visée avec le déclencheur |
| Séparation anatomique | Généralement subtile et naturelle | Habituellement plus nette, même si le périnée peut parfois paraître trop court ou renforcé |
| Préservation de la scène et du style | Préserve plus régulièrement la composition et le style d’origine | Peut attirer davantage l’attention vers la cible anatomique |
| Usage recommandé | Assistant anatomique transparent aux côtés d’autres LoRA NSFW | Interrupteur explicite et correction plus forte et ciblée |
| Élément | Valeur |
|---|---|
| Modèle de base | Qwen Image 2512 |
| Images d’entraînement | 1 046 |
| GPU | NVIDIA RTX PRO 6000 Blackwell / 96 GB |
| Rank | 32 |
| Taille du lot | 8 |
| Étapes d’entraînement | 8 000 |
| Déclencheur | zl_female_anatomy_fix |
zl_female_anatomy_fix près du début du prompt positif0.6-0.9Ce LoRA n’est pas recommandé comme unique adapter actif. Utilisé seul, il peut influencer l’anatomie explicite, mais il ne fournit pas l’identité générale, le style, l’environnement ou la distribution photographique attendus d’un LoRA principal.
Chaque exemple utilise la même description des deux côtés. Le résultat Triggered v2 ajoute zl_female_anatomy_fix au début du prompt.
Sans le LoRA de correction anatomique

Avec Triggered v2

Prompt
zl_female_anatomy_fix, a clearly adult athletic woman running nude except for trail-running shoes, moving uphill away from the camera on a sunny mountain trail, extreme ground-level rear view from a camera only 15 cm above the trail and positioned directly behind her, wide-angle head-to-shoes full-body composition, one foot lifted during a long natural stride, legs clearly separated, camera looking upward beneath the pelvis, buttocks and thighs not blocking the anatomical region, clearly visible anus, coherent perineum and natural vulva as three distinct continuous structures, realistic adult proportions, bright sunlight beneath a vivid blue sky filled with large sculpted white clouds, dramatic outdoor sports photography
| Sans le LoRA de correction anatomique | Avec Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman posing on all fours indoors, looking back over her shoulder, low rear view, white thigh garters, legs apart, clearly visible natural vulva and anus with a coherent perineum, soft daylight, realistic glamour photography
| Sans le LoRA de correction anatomique | Avec Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman kneeling forward on a white bed, close low rear view, legs apart, clearly visible natural vulva and anus with correct separation, bright bedroom daylight, realistic photography
| Sans le LoRA de correction anatomique | Avec Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman kneeling on all fours at the edge of a beach, looking back over her shoulder, low rear three-quarter view, legs apart, clearly visible natural vulva and anus with a coherent perineum, wet skin with fine sand and water droplets, dramatic overcast daylight, realistic photography
Ne jugez pas un LoRA sur une seule graine héroïque. Quand je cours, je ne mentionne pas uniquement le kilomètre en descente avec vent dans le dos en omettant discrètement les côtes. Testez plusieurs graines, poids, morphologies, distances et angles de caméra. Une excellente image est un échantillon ; la régularité est la véritable ligne d’arrivée.
Les fichiers téléchargeables reference.py et reference_low_vram.py chargent Qwen-Image-2512 en BF16, fusionnent l’adapter Lightning communautaire et appliquent Triggered v2. Un GPU CUDA doté d’une grande quantité de mémoire est recommandé pour la version standard.
pip install -U diffusers transformers accelerate peft safetensors
<details>
<summary><strong>▶ Afficher le code Python standard</strong></summary>
"""Qwen Image 2512 + Triggered Female Anatomy Fix LoRA v2."""
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-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2"
)
LORA_WEIGHT_NAME = "anusv2trigger.safetensors"
DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": math.log(3),
"invert_sigmas": False,
"max_image_seq_len": 8192,
"max_shift": math.log(3),
"num_train_timesteps": 1000,
"shift": 1.0,
"shift_terminal": None,
"stochastic_sampling": False,
"time_shift_type": "exponential",
"use_beta_sigmas": False,
"use_dynamic_shifting": True,
"use_exponential_sigmas": False,
"use_karras_sigmas": False,
}
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipe = DiffusionPipeline.from_pretrained(
BASE_MODEL_ID,
scheduler=scheduler,
torch_dtype=DTYPE,
).to(DEVICE)
pipe.load_lora_weights(
LIGHTNING_MODEL_ID,
weight_name=LIGHTNING_WEIGHT_NAME,
adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()
pipe.load_lora_weights(
LORA_MODEL_ID,
weight_name=LORA_WEIGHT_NAME,
adapter_name="anatomy_fix_triggered_v2",
)
pipe.set_adapters(
["anatomy_fix_triggered_v2"],
adapter_weights=[LORA_SCALE],
)
prompt = (
"zl_female_anatomy_fix, a realistic photograph of a clearly adult woman, "
"anatomically coherent external female anatomy, natural pelvic structure, "
"realistic 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("female_anatomy_fix_triggered_v2.png")
</details>
<details>
<summary><strong>▶ Afficher le code Python pour faible VRAM</strong></summary>
"""Low-VRAM Qwen Image 2512 + Triggered Female Anatomy Fix LoRA v2."""
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-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2"
)
LORA_WEIGHT_NAME = "anusv2trigger.safetensors"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": math.log(3),
"invert_sigmas": False,
"max_image_seq_len": 8192,
"max_shift": math.log(3),
"num_train_timesteps": 1000,
"shift": 1.0,
"shift_terminal": None,
"stochastic_sampling": False,
"time_shift_type": "exponential",
"use_beta_sigmas": False,
"use_dynamic_shifting": True,
"use_exponential_sigmas": False,
"use_karras_sigmas": False,
}
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipe = DiffusionPipeline.from_pretrained(
BASE_MODEL_ID,
scheduler=scheduler,
torch_dtype=DTYPE,
low_cpu_mem_usage=True,
)
pipe.load_lora_weights(
LIGHTNING_MODEL_ID,
weight_name=LIGHTNING_WEIGHT_NAME,
adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()
pipe.load_lora_weights(
LORA_MODEL_ID,
weight_name=LORA_WEIGHT_NAME,
adapter_name="anatomy_fix_triggered_v2",
)
pipe.set_adapters(
["anatomy_fix_triggered_v2"],
adapter_weights=[LORA_SCALE],
)
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()
prompt = (
"zl_female_anatomy_fix, a realistic photograph of a clearly adult woman, "
"anatomically coherent external female anatomy, natural pelvic structure, "
"realistic 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("female_anatomy_fix_triggered_v2_low_vram.png")
</details>
Ce modèle est destiné exclusivement à des créations adultes, légales et consensuelles. Ne l’utilisez pas pour :
Il appartient aux utilisateurs de vérifier que leurs prompts, leurs sources et leurs contenus générés respectent les lois et les politiques applicables.
Ce LoRA peut être utilisé uniquement à des fins personnelles et non commerciales. Consultez le fichier LICENSE du dépôt pour connaître les conditions complètes.
Pour toute demande de licence commerciale, veuillez contacter l’auteur.
Le modèle annoncé ici est maintenant publié : Qwen Image Male Anatomy Realism - NSFW LoRA v1. Le texte et l’image ci-dessous sont conservés comme aperçu historique de son développement.
Après avoir terminé un côté de l’audit anatomique, il serait étrangement commode de déclarer le contrôle qualité achevé. Le prochain projet tourne donc le bloc-notes vers l’anatomie masculine adulte : gland, hampe, scrotum, attache pubienne, périnée, et le concept apparemment avancé selon lequel deux testicules doivent rester deux structures distinctes — et non devenir un pouf avec une couture centrale dessinée de mémoire.
Les modèles existants oscillent souvent entre minimalisme extrême et surcompensation héroïque. Un résultat semble assemblé avec un accessoire de mannequin et de l’optimisme ; le suivant paraît conçu pour une espèce vivant sous une autre gravité. Aucun des deux ne décrit utilement un corps humain ordinaire. En plus de fournir aux femmes une fiche technique délicieusement inexacte, ces fantasmes statistiques peuvent pousser des hommes parfaitement normaux à subir, face au miroir, un entretien d’évaluation mené par leur propre estime de soi. L’objectif est plus simple : une anatomie cohérente, connectée et reconnaissablement humaine, dans des dimensions qui ne réclament pas un ingénieur structure.
Il s’agit d’un aperçu préliminaire provenant d’un travail séparé en cours, et non d’une image produite par ce Female Anatomy Fix LoRA. Le sujet représenté est adulte.
<details> <summary><strong>🔞 Cliquer pour révéler un aperçu explicite de l’anatomie masculine adulteL’image suivante est une étude anatomique directe d’un homme adulte. Elle est masquée par défaut afin que son affichage résulte d’un choix délibéré plutôt que d’un changement de décor inattendu. </strong></summary>

Une précision physiologique avant que quelqu’un ne confonde cet aperçu avec de la médecine sportive : les hommes ne courent normalement pas sur des sentiers de montagne exigeants en pleine érection. Si cela vous arrive pendant votre entraînement habituel en côte, ajuster le poids du LoRA n’est probablement pas la première mesure à prendre. Cette condition volontairement improbable sert de test de résistance anatomique : mouvement, angle latéral, tension musculaire, attache pubienne, hampe, gland, scrotum et deux testicules distincts doivent rester cohérents en même temps. C’est la conception d’un benchmark, pas un conseil de course.
</details> </details> <details name="language-version"> <summary><strong>🌎 Español latinoamericano — mostrar la versión completa</strong></summary>
Advertencia de contenido para adultos: este repositorio contiene imágenes explícitas generadas por IA y está dirigido exclusivamente a personas adultas. Todas las personas descritas en los ejemplos y prompts de referencia son adultas.
Muchos LoRA de texto a imagen pueden reproducir los reflejos cromados de una motocicleta, cada grano de arena y el clima emocional exacto de una tormenta que se aproxima; luego llegan a la pelvis y entregan un marcador de posición. Este LoRA fue entrenado para reparar esa carencia concreta en generaciones que, por lo demás, son impresionantes.
Entrenar LoRA se parece sorprendentemente a correr largas distancias. Entreno modelos, pero también corro alrededor de 160 millas al mes: unas 5,3 millas, o aproximadamente 8.500-11.500 pasos humanos al día según la zancada. Mis piernas pueden registrar en un día normal más pasos que este LoRA durante sus 8.000 pasos de entrenamiento. Sin embargo, ni un corredor ni un modelo mejoran por acumular pasos. La fatiga revela la mala técnica de un corredor; un modelo puede ocultar la suya detrás de una semilla favorecedora. Por eso la validación importa más que el kilometraje.
Este es un LoRA de asistencia anatómica, no un modelo autónomo completo de personaje, identidad, estilo o composición. Funciona mejor junto a otro LoRA NSFW que ya aporte la persona, el lenguaje visual, la escena o el carácter fotográfico. Piense en el LoRA principal como el arquitecto y en este adaptador como el especialista que pregunta discretamente si el plano de plomería realmente estaba incluido.
¿Recuerda al Pixiu(貔貅), la bestia mitológica que atrae la riqueza y que mencioné en la ficha de Triggerless v2? Tradicionalmente se dice que no tiene ano: excelente para retener tesoros, menos útil como referencia anatómica. Y no, pasar todo el día sentado no crea anos adicionales. Puede, sin embargo, favorecer problemas anorrectales que hacen que el original parezca tener exceso de personal: otra razón para levantarse y salir. Somos humanos, pero seguimos siendo animales moldeados por el movimiento, la luz, la distancia y el terreno. Salga, camine, corra, suba una colina y deje que el cuerpo recuerde aquello que las pantallas —y a veces los modelos— saben olvidar.
El propósito de tener mejores herramientas no es mantenernos más tiempo frente a ellas. Entrene el modelo, inspeccione los detalles, cierre la pantalla y vuelva a la imagen más amplia: un cuerpo que se mueve, un paisaje que no puede comprimirse en un prompt y una vida cuyo valor no puede medirse solo en pasos.
La Triggerless v2 fue diseñada para actuar discretamente mediante lenguaje descriptivo corriente. Es conveniente cuando la corrección debe estar siempre disponible, pero algunos flujos de trabajo necesitan un interruptor más claro: activarla deliberadamente, aumentar su influencia y mantenerla apartada cuando no se desea.
Ese es el propósito de Triggered v2. No sustituye a Triggerless v2 ni se llama v3 porque la palabra activadora haya recibido de pronto un ascenso. Ambos modelos pertenecen a la misma familia v2 validada: uno espera educadamente en segundo plano; el otro trae la caja de herramientas solo cuando se le llama por su nombre.
La palabra activadora obligatoria es:
zl_female_anatomy_fix
| Dimensión | Versión sin activador | Versión con activador |
|---|---|---|
| Apariencia natural | Más orgánica; la corrección se integra naturalmente | Más firme; la zona corregida puede parecer deliberadamente destacada |
| Visibilidad anatómica | Los ángulos difíciles todavía pueden omitir el ano | Es más probable que produzca la anatomía deseada con el activador |
| Separación anatómica | Por lo general sutil y natural | Normalmente más clara, aunque el perineo puede parecer demasiado corto o reforzado |
| Conservación de escena y estilo | Conserva mejor la composición y el estilo originales | Puede atraer más énfasis hacia el objetivo anatómico |
| Uso recomendado | Asistente anatómico transparente junto a otros LoRA NSFW | Interruptor explícito y corrección más intensa y dirigida |
| Elemento | Valor |
|---|---|
| Modelo base | Qwen Image 2512 |
| Imágenes de entrenamiento | 1.046 |
| GPU | NVIDIA RTX PRO 6000 Blackwell / 96 GB |
| Rank | 32 |
| Tamaño del lote | 8 |
| Pasos de entrenamiento | 8.000 |
| Activador | zl_female_anatomy_fix |
zl_female_anatomy_fix cerca del comienzo del prompt positivo0.6-0.9No se recomienda utilizar este LoRA como único adaptador activo. Por sí solo puede influir en la anatomía explícita, pero no aporta la identidad, el estilo, el entorno ni la distribución fotográfica esperados de un LoRA principal.
Cada ejemplo utiliza la misma descripción en ambos lados. Triggered v2 añade zl_female_anatomy_fix al principio del prompt.
Sin el LoRA de corrección anatómica

Con Triggered v2

Prompt
zl_female_anatomy_fix, a clearly adult athletic woman running nude except for trail-running shoes, moving uphill away from the camera on a sunny mountain trail, extreme ground-level rear view from a camera only 15 cm above the trail and positioned directly behind her, wide-angle head-to-shoes full-body composition, one foot lifted during a long natural stride, legs clearly separated, camera looking upward beneath the pelvis, buttocks and thighs not blocking the anatomical region, clearly visible anus, coherent perineum and natural vulva as three distinct continuous structures, realistic adult proportions, bright sunlight beneath a vivid blue sky filled with large sculpted white clouds, dramatic outdoor sports photography
| Sin el LoRA de corrección anatómica | Con Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman posing on all fours indoors, looking back over her shoulder, low rear view, white thigh garters, legs apart, clearly visible natural vulva and anus with a coherent perineum, soft daylight, realistic glamour photography
| Sin el LoRA de corrección anatómica | Con Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman kneeling forward on a white bed, close low rear view, legs apart, clearly visible natural vulva and anus with correct separation, bright bedroom daylight, realistic photography
| Sin el LoRA de corrección anatómica | Con Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman kneeling on all fours at the edge of a beach, looking back over her shoulder, low rear three-quarter view, legs apart, clearly visible natural vulva and anus with a coherent perineum, wet skin with fine sand and water droplets, dramatic overcast daylight, realistic photography
No juzgue un LoRA por una sola semilla heroica. Cuando corro, no informo únicamente de la milla cuesta abajo con viento de cola mientras omito discretamente las subidas. Pruebe varias semillas, pesos, tipos de cuerpo, distancias y ángulos de cámara. Una imagen excelente es una muestra; la consistencia es la verdadera línea de meta.
Los archivos descargables reference.py y reference_low_vram.py cargan Qwen-Image-2512 en BF16, fusionan el adaptador Lightning comunitario y aplican Triggered v2. Para la versión estándar se recomienda una GPU CUDA con abundante memoria.
pip install -U diffusers transformers accelerate peft safetensors
<details>
<summary><strong>▶ Mostrar referencia Python estándar</strong></summary>
"""Qwen Image 2512 + Triggered Female Anatomy Fix LoRA v2."""
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-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2"
)
LORA_WEIGHT_NAME = "anusv2trigger.safetensors"
DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": math.log(3),
"invert_sigmas": False,
"max_image_seq_len": 8192,
"max_shift": math.log(3),
"num_train_timesteps": 1000,
"shift": 1.0,
"shift_terminal": None,
"stochastic_sampling": False,
"time_shift_type": "exponential",
"use_beta_sigmas": False,
"use_dynamic_shifting": True,
"use_exponential_sigmas": False,
"use_karras_sigmas": False,
}
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipe = DiffusionPipeline.from_pretrained(
BASE_MODEL_ID,
scheduler=scheduler,
torch_dtype=DTYPE,
).to(DEVICE)
pipe.load_lora_weights(
LIGHTNING_MODEL_ID,
weight_name=LIGHTNING_WEIGHT_NAME,
adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()
pipe.load_lora_weights(
LORA_MODEL_ID,
weight_name=LORA_WEIGHT_NAME,
adapter_name="anatomy_fix_triggered_v2",
)
pipe.set_adapters(
["anatomy_fix_triggered_v2"],
adapter_weights=[LORA_SCALE],
)
prompt = (
"zl_female_anatomy_fix, a realistic photograph of a clearly adult woman, "
"anatomically coherent external female anatomy, natural pelvic structure, "
"realistic 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("female_anatomy_fix_triggered_v2.png")
</details>
<details>
<summary><strong>▶ Mostrar referencia Python para poca VRAM</strong></summary>
"""Low-VRAM Qwen Image 2512 + Triggered Female Anatomy Fix LoRA v2."""
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-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2"
)
LORA_WEIGHT_NAME = "anusv2trigger.safetensors"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": math.log(3),
"invert_sigmas": False,
"max_image_seq_len": 8192,
"max_shift": math.log(3),
"num_train_timesteps": 1000,
"shift": 1.0,
"shift_terminal": None,
"stochastic_sampling": False,
"time_shift_type": "exponential",
"use_beta_sigmas": False,
"use_dynamic_shifting": True,
"use_exponential_sigmas": False,
"use_karras_sigmas": False,
}
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipe = DiffusionPipeline.from_pretrained(
BASE_MODEL_ID,
scheduler=scheduler,
torch_dtype=DTYPE,
low_cpu_mem_usage=True,
)
pipe.load_lora_weights(
LIGHTNING_MODEL_ID,
weight_name=LIGHTNING_WEIGHT_NAME,
adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()
pipe.load_lora_weights(
LORA_MODEL_ID,
weight_name=LORA_WEIGHT_NAME,
adapter_name="anatomy_fix_triggered_v2",
)
pipe.set_adapters(
["anatomy_fix_triggered_v2"],
adapter_weights=[LORA_SCALE],
)
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()
prompt = (
"zl_female_anatomy_fix, a realistic photograph of a clearly adult woman, "
"anatomically coherent external female anatomy, natural pelvic structure, "
"realistic 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("female_anatomy_fix_triggered_v2_low_vram.png")
</details>
Este modelo está destinado exclusivamente a trabajos creativos adultos, legales y consensuados. No lo utilice para:
Los usuarios deben garantizar que sus prompts, materiales de origen y contenidos generados cumplan las leyes y políticas aplicables.
Este LoRA solo puede utilizarse con fines personales y no comerciales. Consulte el archivo LICENSE para conocer las condiciones completas.
Para consultas sobre licencias comerciales, póngase en contacto con el autor.
El modelo anunciado aquí ya está publicado: Qwen Image Male Anatomy Realism - NSFW LoRA v1. El texto y la imagen siguientes se conservan como avance histórico de su desarrollo.
Tras completar un lado de la auditoría anatómica, sería sospechosamente cómodo declarar terminado el control de calidad. El próximo proyecto dirige la carpeta hacia la anatomía masculina adulta: glande, cuerpo del pene, escroto, unión púbica, perineo y el concepto aparentemente avanzado de que dos testículos deben seguir siendo dos estructuras distinguibles, no convertirse en un puf con una costura central dibujada de memoria.
Los modelos actuales suelen alternar entre minimalismo extremo y sobrecompensación heroica. Un resultado parece ensamblado con un accesorio de maniquí y optimismo; el siguiente parece diseñado para una especie que vive bajo otra gravedad. Ninguno describe de forma útil un cuerpo humano corriente. Además de entregar a las mujeres una ficha técnica maravillosamente inexacta, estas fantasías pueden enviar a hombres perfectamente normales a una evaluación de desempeño dirigida por su propia autoestima frente al espejo. El objetivo es una anatomía coherente, conectada y reconociblemente humana, con dimensiones que no exijan un ingeniero estructural.
Este es un avance preliminar de un trabajo independiente en curso, no una imagen producida por este Female Anatomy Fix LoRA. El sujeto representado es adulto.
<details> <summary><strong>🔞 Haga clic para mostrar un avance explícito de anatomía masculina adultaLa imagen siguiente es un estudio anatómico directo de un hombre adulto. Está oculta de forma predeterminada para que verla sea una elección deliberada y no un cambio de paisaje inesperado. </strong></summary>

Una nota fisiológica antes de que alguien confunda este avance con medicina deportiva: los hombres no suelen correr por senderos de montaña con una erección completa. Si eso sucede durante su entrenamiento habitual en cuesta, ajustar el peso del LoRA probablemente no sea el primer paso para resolverlo. La condición deliberadamente improbable funciona como prueba de estrés anatómica: movimiento, ángulo lateral, tensión muscular, unión púbica, cuerpo del pene, glande, escroto y dos testículos distintos deben conservar la coherencia al mismo tiempo. Esto es diseño de benchmarks, no consejos para correr.
</details> </details>
Adult content warning: This repository contains explicit AI-generated imagery and is intended only for adults. Every person described in the examples and reference prompts is an adult.
Many text-to-image LoRAs can render chrome reflections on a motorcycle, individual grains of sand, and the precise emotional weather of an approaching storm - then reach the pelvis and submit a placeholder. This LoRA was trained to help repair that particular gap in otherwise impressive generations.
Training LoRAs is surprisingly similar to distance running. I train models, but I also run roughly 160 miles every month - about 5.3 miles, or roughly 8,500-11,500 human steps, every day depending on stride - which has taught me that mileage is not the same thing as form. In other words, my legs may log more steps in an ordinary day than this LoRA completed during its entire 8,000-step training run. Yet neither a runner nor a model improves merely by accumulating steps. Running eventually exposes poor form through fatigue; a model can hide poor form behind one flattering seed. That is why validation matters more than mileage.
This is an anatomy-assist LoRA, not a complete standalone character, identity, style, or composition model. It works best beside another NSFW LoRA that already supplies the person, visual language, scene, or photographic character. Think of the primary LoRA as the architect and this adapter as the specialist who quietly asks whether the plumbing plan was actually included.
Remember the Pixiu(貔貅), the mythical wealth-attracting beast I mentioned in the Triggerless v2 model card? It is traditionally said to have no anus - excellent for retaining treasure, less useful as a default anatomy reference. And no, sitting all day does not create extra anuses. It may, however, encourage the sort of anorectal trouble that makes the original one feel overstaffed - another reason to stand up and go outside. We are human, but we are still animals shaped by movement, daylight, distance, and terrain. Keep something elemental in your life: go outside, hike, run, climb a hill, and let the body remember what screens - and sometimes models - are remarkably good at forgetting.
The purpose of better tools is not to keep us in front of them longer. Train the model, inspect the details, then close the screen and return to the larger image: a body that moves, a landscape that cannot be compressed into a prompt, and a life whose value cannot be measured in steps alone.
The Triggerless v2 was designed to work quietly through ordinary descriptive language. That is convenient when anatomy correction should remain continuously available, but some workflows need a clearer switch: activate the correction deliberately, raise its influence, and keep it out of the way when it is not wanted.
That is the purpose of Triggered v2. It is not a replacement for Triggerless v2 and it is not called v3 because the trigger word did not suddenly earn a promotion. Both models belong to the same validated v2 family; one waits politely in the background, while the other brings the toolbox only when called by name.
The required trigger word is:
zl_female_anatomy_fix
| Dimension | Triggerless Version | Trigger-Based Version |
|---|---|---|
| Natural appearance | More organic; the corrected anatomy feels naturally integrated into the original image | More assertive; the corrected area can occasionally look deliberately emphasized |
| Anatomical visibility | Difficult angles may still occasionally omit the anus | More likely to produce the intended anatomy when the trigger word is used |
| Anatomical separation | Generally subtle and natural | Usually clearer, although the perineum may sometimes appear too short or overly reinforced |
| Scene and style preservation | More consistent at preserving the original composition and visual style | More likely to pull the visual emphasis toward the anatomical target |
| Recommended use | Best used as a seamless anatomical assistant alongside other NSFW LoRAs | Best for users who want an explicit on/off switch and stronger, targeted correction |
| Item | Value |
|---|---|
| Base model | Qwen Image 2512 |
| Training images | 1,046 |
| GPU | NVIDIA RTX PRO 6000 Blackwell / 96 GB |
| Rank | 32 |
| Batch size | 8 |
| Training steps | 8,000 |
| Trigger | zl_female_anatomy_fix |
zl_female_anatomy_fix near the beginning of the positive prompt0.6-0.9This LoRA is not recommended as the only active adapter. Used alone, it may influence explicit anatomy, but it does not provide the broader identity, styling, environment, or photographic distribution expected from a primary LoRA.
Each example uses the same descriptive setup on both sides. The Triggered v2 result adds the trigger word zl_female_anatomy_fix at the beginning of the prompt.
Without anatomy-fix LoRA

With Triggered v2

Prompt
zl_female_anatomy_fix, a clearly adult athletic woman running nude except for trail-running shoes, moving uphill away from the camera on a sunny mountain trail, extreme ground-level rear view from a camera only 15 cm above the trail and positioned directly behind her, wide-angle head-to-shoes full-body composition, one foot lifted during a long natural stride, legs clearly separated, camera looking upward beneath the pelvis, buttocks and thighs not blocking the anatomical region, clearly visible anus, coherent perineum and natural vulva as three distinct continuous structures, realistic adult proportions, bright sunlight beneath a vivid blue sky filled with large sculpted white clouds, dramatic outdoor sports photography
| Without anatomy-fix LoRA | With Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman posing on all fours indoors, looking back over her shoulder, low rear view, white thigh garters, legs apart, clearly visible natural vulva and anus with a coherent perineum, soft daylight, realistic glamour photography
| Without anatomy-fix LoRA | With Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman kneeling forward on a white bed, close low rear view, legs apart, clearly visible natural vulva and anus with correct separation, bright bedroom daylight, realistic photography
| Without anatomy-fix LoRA | With Triggered v2 |
|---|---|
![]() | ![]() |
Prompt
zl_female_anatomy_fix, a clearly adult East Asian woman kneeling on all fours at the edge of a beach, looking back over her shoulder, low rear three-quarter view, legs apart, clearly visible natural vulva and anus with a coherent perineum, wet skin with fine sand and water droplets, dramatic overcast daylight, realistic photography
Do not judge a LoRA from one heroic seed. When I run, I do not report only the downhill mile with a tailwind and quietly omit the hills. Test several seeds, weights, body types, distances, and camera angles before declaring victory. One excellent image is a sample; consistency is the actual finish line.
The downloadable reference.py and reference_low_vram.py load Qwen-Image-2512 in BF16, fuse the community Lightning adapter, and apply Triggered v2. A high-memory CUDA GPU is recommended for the standard version.
pip install -U diffusers transformers accelerate peft safetensors
<details>
<summary><strong>▶ Show Standard Python Reference</strong></summary>
"""Qwen Image 2512 + Triggered Female Anatomy Fix LoRA v2."""
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-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2"
)
LORA_WEIGHT_NAME = "anusv2trigger.safetensors"
DEVICE = "cuda"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": math.log(3),
"invert_sigmas": False,
"max_image_seq_len": 8192,
"max_shift": math.log(3),
"num_train_timesteps": 1000,
"shift": 1.0,
"shift_terminal": None,
"stochastic_sampling": False,
"time_shift_type": "exponential",
"use_beta_sigmas": False,
"use_dynamic_shifting": True,
"use_exponential_sigmas": False,
"use_karras_sigmas": False,
}
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipe = DiffusionPipeline.from_pretrained(
BASE_MODEL_ID,
scheduler=scheduler,
torch_dtype=DTYPE,
).to(DEVICE)
pipe.load_lora_weights(
LIGHTNING_MODEL_ID,
weight_name=LIGHTNING_WEIGHT_NAME,
adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()
pipe.load_lora_weights(
LORA_MODEL_ID,
weight_name=LORA_WEIGHT_NAME,
adapter_name="anatomy_fix_triggered_v2",
)
pipe.set_adapters(
["anatomy_fix_triggered_v2"],
adapter_weights=[LORA_SCALE],
)
prompt = (
"zl_female_anatomy_fix, a realistic photograph of a clearly adult woman, "
"anatomically coherent external female anatomy, natural pelvic structure, "
"realistic 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("female_anatomy_fix_triggered_v2.png")
</details>
<details>
<summary><strong>▶ Show Low-VRAM Python Reference</strong></summary>
"""Low-VRAM Qwen Image 2512 + Triggered Female Anatomy Fix LoRA v2."""
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-Female-Anatomy-Fix-NSFW-LoRA-Triggered-v2"
)
LORA_WEIGHT_NAME = "anusv2trigger.safetensors"
DTYPE = torch.bfloat16
LORA_SCALE = 0.75
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": math.log(3),
"invert_sigmas": False,
"max_image_seq_len": 8192,
"max_shift": math.log(3),
"num_train_timesteps": 1000,
"shift": 1.0,
"shift_terminal": None,
"stochastic_sampling": False,
"time_shift_type": "exponential",
"use_beta_sigmas": False,
"use_dynamic_shifting": True,
"use_exponential_sigmas": False,
"use_karras_sigmas": False,
}
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipe = DiffusionPipeline.from_pretrained(
BASE_MODEL_ID,
scheduler=scheduler,
torch_dtype=DTYPE,
low_cpu_mem_usage=True,
)
pipe.load_lora_weights(
LIGHTNING_MODEL_ID,
weight_name=LIGHTNING_WEIGHT_NAME,
adapter_name="lightning",
)
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
pipe.unload_lora_weights()
pipe.load_lora_weights(
LORA_MODEL_ID,
weight_name=LORA_WEIGHT_NAME,
adapter_name="anatomy_fix_triggered_v2",
)
pipe.set_adapters(
["anatomy_fix_triggered_v2"],
adapter_weights=[LORA_SCALE],
)
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()
prompt = (
"zl_female_anatomy_fix, a realistic photograph of a clearly adult woman, "
"anatomically coherent external female anatomy, natural pelvic structure, "
"realistic 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("female_anatomy_fix_triggered_v2_low_vram.png")
</details>
This model is intended exclusively for lawful, consensual, adult creative work. Do not use it to:
Users are responsible for ensuring that their prompts, source material, and generated content comply with applicable laws and platform policies.
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.
The model announced here has now been released: Qwen Image Male Anatomy Realism - NSFW LoRA v1. The text and image below are preserved as a historical preview of its development.
Having completed one side of the anatomical audit, it would be suspiciously convenient to declare quality control finished. The next project turns the clipboard toward adult male anatomy: glans, shaft, scrotum, pubic attachment, perineum, and the apparently advanced concept that two testicles should remain two distinguishable structures—not become one beanbag with a center seam drawn from memory.
Existing models often alternate between extreme minimalism and heroic overcompensation. One result looks assembled from a mannequin accessory and optimism; the next appears designed for a species living under different gravity. Neither is a useful description of an ordinary human body. Besides giving women a wonderfully inaccurate specification sheet, these statistical fantasies can send perfectly normal men into a performance review conducted by their own bathroom mirror. The aim of the next LoRA is simpler: coherent, connected, recognizably human anatomy at dimensions that do not require a structural engineer.
This is an early preview from a separate work in progress, not an output from this Female Anatomy Fix LoRA. The depicted subject is an adult.
<details> <summary><strong>🔞 Click to reveal an explicit adult male anatomy previewThe following image is a direct adult male anatomical study. It is collapsed by default so that viewing it is a deliberate choice rather than an unexpected change of scenery. </strong></summary>

One physiological footnote before anyone mistakes this preview for sports medicine: men do not ordinarily run hard mountain trails with a full erection. If that happens during your regular hill workout, adjusting the LoRA weight is probably not the first troubleshooting step. The deliberately unlikely condition is used here as an anatomical stress test—motion, side angle, muscular tension, pubic attachment, shaft, glans, scrotum, and two distinct testicles must all remain coherent at once. This is benchmark design, not running advice.
</details>The earlier Eastern European Female Anatomy NSFW LoRA v1 explored a broad Eastern European photographic distribution across body types, poses, lighting, environments, and anatomy. This new project moves the camera west—not because latent space recognizes borders, and certainly not because a seed has ever produced a valid passport photo. It explores a separate Western European / American distribution with different faces, body types, styling, locations, and photographic atmosphere.
The two models are companions, not a national beauty ranking. Geography is already complicated enough without asking the sampler to clear customs; the useful comparison is diversity, controllability, and whether a model can render more than one woman wearing the same borrowed face.
This is an early preview from another model still in training, not an output from this Female Anatomy Fix LoRA. The depicted subject is an adult, and the sample was generated without the anatomy-fix adapter.
<details> <summary><strong>🔞 Click to reveal an explicit adult preview from the upcoming Western European / American model</strong></summary>