Downloads · 30 days
2.1K
32% of all-time downloads
perpetual3x/Eastern-European-Female-Anatomy-NSFW-LoRA-v2
Eastern-European-Female-Anatomy-NSFW-LoRA-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:flex-start; overflow-x:auto;"<img alt="Visitors" src="https://visitor-badge.laobi.icu/badge?pageid=perpetual3x.Eastern-European-Female-Anatomy-NSFW-LoRA…
Downloads · 30 days
2.1K
32% of all-time downloads
All-time downloads
6.5K
Public
Repo size
1.6 GB
Likes
11
Public
Click a slice to open those files.
.safetensors1.5 GB · 97%
From the Hugging Face model README
🇬🇧 English - displayed below

Pour le projet d'origine, voir l'édition v1. La v2 utilise un jeu d'entraînement entièrement nouveau : il ne s'agit pas des images de la v1 sous une autre étiquette.
Je buvais un café à Fairbanks, en Alaska, quand Qwen Image 2.1 est arrivé. La lumière du Nord prenait son temps ; mon GPU y a vu une invitation à faire des heures supplémentaires. J'ai donc entraîné ce nouveau jeu de données deux fois : sur Qwen Image 2512 et sur Qwen Image 2.1. L'image suivante illustre le café ; ce n'est pas une photographie documentaire de l'instant de sortie.

Qwen Image 2.1 sait déjà produire des images NSFW. Mais accepter de dessiner une personne dévêtue et savoir dessiner une anatomie cohérente sont deux qualifications différentes. La première permet d'entrer ; pour la seconde, il faut encore lire le plan. Ce LoRA tente d'améliorer ce deuxième point sans prétendre que l'un ou l'autre modèle de base a réussi tous les examens.
L'entraînement de la version 2.1 pourrait être insuffisant. Je pourrai le poursuivre après des tests plus larges ; 10 000 étapes ne sont pas déclarées optimales pour tous les cas.
person_0001 à person_0123.Il s'agit d'impressions qualitatives, pas d'une promesse pour chaque seed ou prompt. Le tableau de la préversion est conservé.
| Aspect | Qwen Image Base 2512 | Qwen Image 2.1 |
|---|---|---|
| Aspect général | Net, poli, idéalisé | Naturel, photographique, moins standardisé par la retouche |
| Peau | Lisse et contrôlée | Plus de texture, de variation et d'imperfections crédibles |
| Corps | Souvent élégant et symétrique | Souvent plus convaincant pour le poids, les tissus et la continuité structurelle |
| Environnement | Séduisant mais parfois mis en scène | Meilleure intégration du sujet, de la lumière, de l'usure et de l'espace |
| Usage privilégié | Présentation raffinée et beauté conventionnelle | Réalisme, caractère documentaire et travail anatomique |
A correspond au LoRA entraîné sur Qwen Image 2512 ; B, à celui entraîné sur Qwen Image 2.1. Ces images sont rapprochées pour l'observation, sans constituer des essais parfaitement identiques en termes de prompt.
A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

| Élément | Valeur |
|---|---|
| Images d'entraînement | 17 050 ; jeu distinct de la v1 |
| GPU | NVIDIA RTX PRO 6000 Blackwell / 96 GB |
| Rank | 64 |
| Taille de lot | 8 |
| Étapes | 10 000 pour chaque version |
| Base A | Qwen Image 2512 |
| Base B | Qwen Image 2.1 |
| Déclenchement | Langage naturel ou tokens de personne/détail appropriés |
| Fichier | Modèle de base requis | Usage |
|---|---|---|
easternfemale_v2512.safetensors | Qwen Image 2512 uniquement | Adaptateur entraîné pour la base 2512 ; utiliser les références d'inférence 2512. |
easternfemale_v2.1.safetensors | Qwen Image 2.1 uniquement | Adaptateur distinct entraîné pour la base 2.1 ; les exemples B suivants utilisent le workflow 2.1 accéléré par Viggle. |
Les deux fichiers de poids sont disponibles dans ce dépôt. Utilisez chacun uniquement avec son modèle de base correspondant.
Pour une personne donnée, le token d'identité peut être complété par des tokens de détail : person_0001, person_0001_vulva, person_0001_anus, person_0001_breasts et person_0001_nipples. Le même schéma va jusqu'à person_0123. Un token de détail demande une association apprise ; il ne garantit ni visibilité ni exactitude indépendamment de la pose. Évitez de mélanger les identifiants de plusieurs personnes si la fidélité d'identité compte.
Le fichier de prompts/captions d'entraînement peut être téléchargé pour inspection ; il n'est pas nécessaire de recopier tous ses tokens lors de l'inférence.
Les exemples B5 à B8 utilisent l'adaptateur Qwen Image 2.1 avec l'accélération Viggle. Les prompts exacts restent en anglais pour pouvoir être réutilisés directement.
| Image générée |
|---|
![]() |
Prompt
A softly textured analog photograph captures an adult nude woman whose presence feels candid and unperformed walking barefoot along an empty corridor, caught between one step and the next inside an old railway waiting room with wooden benches and frosted windows. A few damp strands of hair rest naturally against her face and shoulders. A gently compressed telephoto view layers the figure, architecture and distant space. Warm practical light mixes gently with cool ambient light from outside. The photograph combines classical figure-study discipline with an unpolished documentary surface. The emotional tone is reflective and slightly melancholic. Keep the gesture physically achievable and the relationship between body, gravity and support convincing.
| Image générée |
|---|
![]() |
Prompt
An adult nude woman is performing a gymnastic squat turn with heels grounded and knees open in the foreground. A detailed, realistic depiction of the interior of an abandoned building, likely a school or public facility. The scene is dimly lit, with light filtering in through large, multi-paned windows that span the length of the room. The windows are mostly intact but show signs of wear and neglect, with some panes appearing dirty or dusty. The windows are framed by dark, possibly metal or wooden, frames that are slightly rusted or weathered. full-body frontal view from pelvis height using a normal lens, explicit clinical fine-art athletic anatomy reference emphasizing a clearly resolved vulva, distinct natural labia, clitoral hood, pubic mound, perineum, coherent inner-thigh transitions and realistic soft-tissue response to the athletic pose; accurate spine, shoulders, joints, hands and feet, realistic balance and muscle tension, soft overcast environmental light, entire body visible, no censoring, no fused anatomy, no extra or duplicated limbs.
| Image générée |
|---|
![]() |
Prompt
A clearly adult woman with a natural athletic build and unretouched skin, wearing a formal embroidered uchikake arranged behind her like a theatrical backdrop rather than covering the figure, and kneeling with the knees separated and the pelvis relaxed in a secluded daimyo garden with a pond, stone bridge and carefully trained pines. a close but contextual frontal composition includes the torso, both thighs and the full pelvis; the clitoral hood and surrounding labial anatomy are clearly readable without exaggeration, obstruction, duplicated tissue or artificial symmetry. warm late-afternoon light across the tatami, documentary-style medium-format photography; wooden geta rest in the background as a quiet cultural detail. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.
Cette scène d'inspiration japonaise teste la généralisation hors du contexte est-européen ; elle ne définit pas l'identité régionale par défaut.
| Image générée |
|---|
![]() |
Prompt
adult woman, full-body nude, anatomically detailed breasts and vulva, explicit anatomical cues to anus under tasteful realism, 18th-century European oil portrait tradition (early Neoclassical), nude, no added garments, deep contrapposto pose with hips angled toward camera and torso twisting, frontal studio-like framing in outdoor scene, camera framed to show lower abdomen to thigh transition with visible anal contour, foggy marsh edge with reed thickets, with reeds moving in humid wind along a river bend, warm late-afternoon sunlight with cool sky bounce, textured oil brush strokes, museum realism, low micro-contrast, cinematic depth, classical composition, outdoor setting, textured brushwork, high detail, museum-grade classical realism.
Les quatre personnes représentées sont adultes. La description de scène reste largement identique ; les tokens de personne et d'anatomie changent. Les résultats dépendent aussi de la seed, du poids du LoRA et du modèle de base.
| Exemple | Image | Tokens du prompt |
|---|---|---|
| P1 | ![]() | person_0040, person_0040_vulva, person_0040_anus |
| P2 | ![]() | person_0069, person_0069_vulva, person_0069_anus |
| P3 | ![]() | person_0101, person_0101_vulva, person_0101_anus |
| P4 | ![]() | person_0007, person_0007_vulva, person_0007_anus |
Prompt commun après les tokens correspondants : nude, adult, light build, front view, close-up, legs spread wide, feet on table, hand on chest, beaded bracelet on left wrist, blonde long hair, smiling, looking at camera, indoor, wooden table, bookshelf in background, soft lighting, studio style. Les prompts complets, copiables individuellement, figurent dans la version anglaise ci-dessous.
| Base et mode | Fichier |
|---|---|
| Qwen Image 2512, mémoire GPU plus élevée | reference_2512.py |
| Qwen Image 2512, transfert séquentiel vers le CPU | reference_2512_low_vram.py |
| Qwen Image 2.1 + Viggle, mémoire GPU plus élevée | reference_21_viggle.py |
| Qwen Image 2.1 + Viggle, mémoire GPU réduite (expérimental) | reference_21_viggle_low_vram.py |
Les deux adaptateurs ne sont pas interchangeables. La variante 2.1 à mémoire vidéo réduite nécessite toujours un GPU CUDA : les composants inactifs restent en RAM et passent sur le GPU au besoin. Elle n'exécute pas toute la génération sur CPU. Ce script n'a pas encore été vérifié avec le poids final ; il peut être lent et demander beaucoup de RAM système.
Ce modèle vise l'anatomie adulte et un réalisme visuel régional ; il ne garantit ni anatomie ni identité parfaites. Les angles difficiles, l'occlusion, les poses extrêmes et les cadrages serrés restent délicats. L'entraînement 2.1 pourra être prolongé si les tests le justifient.
Réservé à la création et à la recherche licites et consenties concernant des adultes clairement identifiés comme tels. Ne l'utilisez pas pour des mineurs, du contenu non consenti, la sexualisation de personnes réelles identifiables sans permission, le harcèlement ou la tromperie.
Utilisation non commerciale uniquement. L'usage commercial ou la redistribution exigent l'autorisation préalable de l'auteur.
</details> <br> <details name="language-version"> <summary><strong>🌐 Español latinoamericano - mostrar la versión completa</strong></summary>
Para conocer el proyecto original, consulta la edición v1. La v2 utiliza un conjunto de entrenamiento completamente nuevo: no son las imágenes de la v1 con otra etiqueta.
Estaba tomando café en Fairbanks, Alaska, cuando apareció Qwen Image 2.1. La luz del norte se tomaba su tiempo; mi GPU lo interpretó como una invitación a trabajar horas extra. Por eso entrené este nuevo conjunto dos veces: con Qwen Image 2512 y con Qwen Image 2.1. La siguiente imagen ilustra la parte del café; no es una fotografía documental del momento del lanzamiento.

Qwen Image 2.1 ya puede generar imágenes NSFW. Sin embargo, estar dispuesto a dibujar a una persona sin ropa y saber dibujar una anatomía coherente son dos cualificaciones distintas. La primera abre la puerta; la segunda exige leer los planos. Este LoRA intenta mejorar la segunda, sin afirmar que ninguno de los modelos base haya aprobado todos los exámenes.
El entrenamiento de la versión 2.1 quizá aún sea insuficiente. Podría continuarlo después de más pruebas; 10 000 pasos no son un óptimo universal.
person_0001 a person_0123.Estas son impresiones cualitativas, no garantías para cada seed o prompt. Se conserva la tabla de la vista previa.
| Aspecto | Qwen Image Base 2512 | Qwen Image 2.1 |
|---|---|---|
| Apariencia general | Limpia, pulida, idealizada | Natural, fotográfica, menos estandarizada por el retoque |
| Piel | Lisa y controlada | Más textura, variación e imperfecciones creíbles |
| Cuerpo | A menudo elegante y simétrico | Suele representar mejor el peso, los tejidos y la continuidad estructural |
| Entorno | Atractivo, aunque a veces escenificado | Mejor integración del sujeto, la luz, el desgaste y el espacio |
| Uso preferente | Presentación refinada y belleza convencional | Realismo, carácter documental y trabajo anatómico |
A corresponde al LoRA entrenado sobre Qwen Image 2512; B, al entrenado sobre Qwen Image 2.1. Las imágenes se presentan juntas para examinarlas, pero no son pruebas idénticas y perfectamente controladas en cuanto al prompt.
A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

| Elemento | Valor |
|---|---|
| Imágenes de entrenamiento | 17 050; conjunto distinto al de la v1 |
| GPU | NVIDIA RTX PRO 6000 Blackwell / 96 GB |
| Rank | 64 |
| Tamaño de lote | 8 |
| Pasos | 10 000 para cada variante |
| Base A | Qwen Image 2512 |
| Base B | Qwen Image 2.1 |
| Activación | Lenguaje natural o tokens correspondientes a persona/detalle |
| Archivo | Modelo base necesario | Uso |
|---|---|---|
easternfemale_v2512.safetensors | Solo Qwen Image 2512 | Adaptador entrenado para 2512; utiliza los ejemplos de inferencia para esa base. |
easternfemale_v2.1.safetensors | Solo Qwen Image 2.1 | Adaptador independiente entrenado para 2.1; los ejemplos B posteriores usan el flujo 2.1 acelerado por Viggle. |
Los dos archivos de pesos están disponibles en este repositorio. Usa cada uno únicamente con su modelo base correspondiente.
Para una persona concreta, el token de identidad puede acompañarse de tokens de detalle: person_0001, person_0001_vulva, person_0001_anus, person_0001_breasts y person_0001_nipples. El mismo patrón llega hasta person_0123. Un token de detalle solicita una asociación aprendida; no garantiza visibilidad ni precisión independientemente de la pose. Evita combinar identificadores de personas diferentes si importa la fidelidad de identidad.
El archivo descargable de prompts/captions de entrenamiento se ofrece para su inspección; no hace falta copiar todos sus tokens al inferir.
Los ejemplos B5 a B8 usan el adaptador Qwen Image 2.1 con aceleración Viggle. Los prompts exactos se conservan en inglés para poder usarlos directamente.
| Imagen generada |
|---|
![]() |
Prompt
A softly textured analog photograph captures an adult nude woman whose presence feels candid and unperformed walking barefoot along an empty corridor, caught between one step and the next inside an old railway waiting room with wooden benches and frosted windows. A few damp strands of hair rest naturally against her face and shoulders. A gently compressed telephoto view layers the figure, architecture and distant space. Warm practical light mixes gently with cool ambient light from outside. The photograph combines classical figure-study discipline with an unpolished documentary surface. The emotional tone is reflective and slightly melancholic. Keep the gesture physically achievable and the relationship between body, gravity and support convincing.
| Imagen generada |
|---|
![]() |
Prompt
An adult nude woman is performing a gymnastic squat turn with heels grounded and knees open in the foreground. A detailed, realistic depiction of the interior of an abandoned building, likely a school or public facility. The scene is dimly lit, with light filtering in through large, multi-paned windows that span the length of the room. The windows are mostly intact but show signs of wear and neglect, with some panes appearing dirty or dusty. The windows are framed by dark, possibly metal or wooden, frames that are slightly rusted or weathered. full-body frontal view from pelvis height using a normal lens, explicit clinical fine-art athletic anatomy reference emphasizing a clearly resolved vulva, distinct natural labia, clitoral hood, pubic mound, perineum, coherent inner-thigh transitions and realistic soft-tissue response to the athletic pose; accurate spine, shoulders, joints, hands and feet, realistic balance and muscle tension, soft overcast environmental light, entire body visible, no censoring, no fused anatomy, no extra or duplicated limbs.
| Imagen generada |
|---|
![]() |
Prompt
A clearly adult woman with a natural athletic build and unretouched skin, wearing a formal embroidered uchikake arranged behind her like a theatrical backdrop rather than covering the figure, and kneeling with the knees separated and the pelvis relaxed in a secluded daimyo garden with a pond, stone bridge and carefully trained pines. a close but contextual frontal composition includes the torso, both thighs and the full pelvis; the clitoral hood and surrounding labial anatomy are clearly readable without exaggeration, obstruction, duplicated tissue or artificial symmetry. warm late-afternoon light across the tatami, documentary-style medium-format photography; wooden geta rest in the background as a quiet cultural detail. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.
Esta escena japonesa comprueba la generalización fuera del enfoque europeo oriental; no define la identidad regional predeterminada.
| Imagen generada |
|---|
![]() |
Prompt
adult woman, full-body nude, anatomically detailed breasts and vulva, explicit anatomical cues to anus under tasteful realism, 18th-century European oil portrait tradition (early Neoclassical), nude, no added garments, deep contrapposto pose with hips angled toward camera and torso twisting, frontal studio-like framing in outdoor scene, camera framed to show lower abdomen to thigh transition with visible anal contour, foggy marsh edge with reed thickets, with reeds moving in humid wind along a river bend, warm late-afternoon sunlight with cool sky bounce, textured oil brush strokes, museum realism, low micro-contrast, cinematic depth, classical composition, outdoor setting, textured brushwork, high detail, museum-grade classical realism.
Las cuatro personas representadas son adultas. La descripción de la escena permanece casi igual, mientras cambian los tokens de persona y anatomía. Los resultados también dependen de la seed, la intensidad del LoRA y el modelo base.
| Ejemplo | Imagen | Tokens del prompt |
|---|---|---|
| P1 | ![]() | person_0040, person_0040_vulva, person_0040_anus |
| P2 | ![]() | person_0069, person_0069_vulva, person_0069_anus |
| P3 | ![]() | person_0101, person_0101_vulva, person_0101_anus |
| P4 | ![]() | person_0007, person_0007_vulva, person_0007_anus |
Prompt compartido tras los tokens correspondientes: nude, adult, light build, front view, close-up, legs spread wide, feet on table, hand on chest, beaded bracelet on left wrist, blonde long hair, smiling, looking at camera, indoor, wooden table, bookshelf in background, soft lighting, studio style. Los prompts completos, listos para copiar por separado, figuran en la versión inglesa más abajo.
| Base y modo | Archivo |
|---|---|
| Qwen Image 2512, mayor memoria de GPU | reference_2512.py |
| Qwen Image 2512, descarga secuencial a CPU | reference_2512_low_vram.py |
| Qwen Image 2.1 + Viggle, mayor memoria de GPU | reference_21_viggle.py |
| Qwen Image 2.1 + Viggle, poca VRAM (experimental) | reference_21_viggle_low_vram.py |
Los adaptadores no son intercambiables. La variante 2.1 de menor VRAM sigue necesitando una GPU CUDA: los componentes inactivos se almacenan en RAM y pasan a la GPU cuando se necesitan. No realiza toda la generación en CPU. Este script todavía no se ha probado con el peso final; puede ser lento y requerir mucha RAM del sistema.
El modelo se centra en anatomía adulta y realismo visual regional; no garantiza anatomía ni identidad perfectas. Los ángulos difíciles, la oclusión, las poses extremas y los encuadres cerrados siguen siendo problemáticos. El entrenamiento 2.1 podría ampliarse si las pruebas lo justifican.
Solo para generación e investigación lícitas y consensuadas con personas claramente adultas. No lo uses para menores, contenido no consensuado, sexualización de personas reales identificables sin permiso, acoso o engaño.
Solo uso no comercial. El uso comercial o la redistribución requieren autorización previa del autor.
</details>
For the original edition and the longer explanation of the project's purpose, see Eastern European Female Anatomy NSFW LoRA v1. Version 2 uses an entirely new training dataset—it is not v1's images with a new label.
I was drinking coffee in Fairbanks, Alaska, when Qwen Image 2.1 appeared. The northern light was generous with its hours; my GPU took that as an invitation to work overtime. So I trained this new dataset twice: once on Qwen Image 2512, and once on Qwen Image 2.1. The image below illustrates the coffee part of the story; it is not a documentary photograph of the release moment.

Qwen Image 2.1 can already generate NSFW imagery. But being willing to draw an unclothed person and being able to draw coherent anatomy are different qualifications. One gets you through the door; the other requires reading the floor plan. This LoRA is an anatomy-focused attempt to improve that second qualification, not a claim that either base model has passed every exam.
The 2.1 training may still be insufficient. I may continue it after broader testing; treat these results as a documented release, not a declaration that 10,000 steps is a universal optimum.
person_0001–person_0123.These are qualitative impressions from the current samples, not a guarantee for every seed or prompt. The comparison table from the preview is retained here.
| Area | Qwen Image Base 2512 | Qwen Image 2.1 |
|---|---|---|
| Overall look | Clean, polished, idealized | Natural, photographic, less cosmetically standardized |
| Skin | Smooth and controlled | More texture, variation, and believable imperfection |
| Body rendering | Often elegant and symmetrical | Usually stronger weight, tissue, and structural continuity |
| Environment | Attractive but occasionally feels staged | Better integration of subject, light, wear, and space |
| Best use | Refined presentation and conventional beauty | Realism, documentary character, and anatomy-oriented work |
A denotes the adapter trained on Qwen Image 2512; B denotes the adapter trained on Qwen Image 2.1. The images below are paired for viewing, but should not be taken as perfectly controlled, prompt-identical experiments.
A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

A — Qwen Image 2512

B — Qwen Image 2.1

| Setting | Value |
|---|---|
| Training images | 17,050; new dataset, distinct from v1 |
| GPU | NVIDIA RTX PRO 6000 Blackwell / 96 GB |
| Rank | 64 |
| Batch size | 8 |
| Training steps | 10,000 per base-model variant |
| Variant A base | Qwen Image 2512 |
| Variant B base | Qwen Image 2.1 |
| Trigger | Natural-language description or the relevant person/detail tokens |
| Weight file | Use with | Purpose |
|---|---|---|
easternfemale_v2512.safetensors | Qwen Image 2512 only | The adapter trained for the 2512 base; use the 2512 inference references below. |
easternfemale_v2.1.safetensors | Qwen Image 2.1 only | The separately trained 2.1 adapter; the later B-series examples use the Viggle-accelerated 2.1 workflow. |
Both weight files are available in this repository. Use each adapter only with its matching base model.
For a person in the training set, use the ID token when identity matters. Detail tokens have the same ID prefix, for example person_0001, person_0001_vulva, person_0001_anus, person_0001_breasts, and person_0001_nipples. The same naming scheme runs through person_0123. A detail token helps request a learned association; it does not guarantee that the feature will be visible, accurate, or independent of pose and context. Do not mix person IDs and detail IDs from different people if identity fidelity matters.
The downloadable training prompt/caption resource contains the source text used for this release. It is supplied for inspection, not as a requirement to copy every token into an inference prompt.
The following B-series examples were generated with the Qwen Image 2.1 adapter using the Viggle acceleration workflow. These are examples, not a claim that the accelerator is identical to unaccelerated inference.

Prompt: A softly textured analog photograph captures an adult nude woman whose presence feels candid and unperformed walking barefoot along an empty corridor, caught between one step and the next inside an old railway waiting room with wooden benches and frosted windows. A few damp strands of hair rest naturally against her face and shoulders. A gently compressed telephoto view layers the figure, architecture and distant space. Warm practical light mixes gently with cool ambient light from outside. The photograph combines classical figure-study discipline with an unpolished documentary surface. The emotional tone is reflective and slightly melancholic. Keep the gesture physically achievable and the relationship between body, gravity and support convincing.

Prompt: An adult nude woman is performing a gymnastic squat turn with heels grounded and knees open in the foreground. A detailed, realistic depiction of the interior of an abandoned building, likely a school or public facility. The scene is dimly lit, with light filtering in through large, multi-paned windows that span the length of the room. The windows are mostly intact but show signs of wear and neglect, with some panes appearing dirty or dusty. The windows are framed by dark, possibly metal or wooden, frames that are slightly rusted or weathered. full-body frontal view from pelvis height using a normal lens, explicit clinical fine-art athletic anatomy reference emphasizing a clearly resolved vulva, distinct natural labia, clitoral hood, pubic mound, perineum, coherent inner-thigh transitions and realistic soft-tissue response to the athletic pose; accurate spine, shoulders, joints, hands and feet, realistic balance and muscle tension, soft overcast environmental light, entire body visible, no censoring, no fused anatomy, no extra or duplicated limbs.

Prompt: A clearly adult woman with a natural athletic build and unretouched skin, wearing a formal embroidered uchikake arranged behind her like a theatrical backdrop rather than covering the figure, and kneeling with the knees separated and the pelvis relaxed in a secluded daimyo garden with a pond, stone bridge and carefully trained pines. a close but contextual frontal composition includes the torso, both thighs and the full pelvis; the clitoral hood and surrounding labial anatomy are clearly readable without exaggeration, obstruction, duplicated tissue or artificial symmetry. warm late-afternoon light across the tatami, documentary-style medium-format photography; wooden geta rest in the background as a quiet cultural detail. Accurate adult proportions, realistic hands and feet, no text, no logo, no watermark.
This Japanese-themed scene tests generalization outside the project's Eastern European focus; it is not presented as the default regional identity.

Prompt: adult woman, full-body nude, anatomically detailed breasts and vulva, explicit anatomical cues to anus under tasteful realism, 18th-century European oil portrait tradition (early Neoclassical), nude, no added garments, deep contrapposto pose with hips angled toward camera and torso twisting, frontal studio-like framing in outdoor scene, camera framed to show lower abdomen to thigh transition with visible anal contour, foggy marsh edge with reed thickets, with reeds moving in humid wind along a river bend, warm late-afternoon sunlight with cool sky bounce, textured oil brush strokes, museum realism, low micro-contrast, cinematic depth, classical composition, outdoor setting, textured brushwork, high detail, museum-grade classical realism.
All four subjects in the following examples are adults. These prompts deliberately hold much of the scene description constant while changing person and matching anatomy tokens. The results still depend on seed, adapter strength, and the base model.
person_0040
Prompt: person_0040, person_0040_vulva, person_0040_anus, nude, adult, light build, front view, close-up, legs spread wide, feet on table, hand on chest, beaded bracelet on left wrist, blonde long hair, smiling, looking at camera, indoor, wooden table, bookshelf in background, soft lighting, studio style.
person_0069
Prompt: person_0069, person_0069_vulva, person_0069_anus, nude, adult, light build, front view, close-up, legs spread wide, feet on table, hand on chest, beaded bracelet on left wrist, blonde long hair, smiling, looking at camera, indoor, wooden table, bookshelf in background, soft lighting, studio style.
person_0101
Prompt: person_0101, person_0101_vulva, person_0101_anus, nude, adult, light build, front view, close-up, legs spread wide, feet on table, hand on chest, beaded bracelet on left wrist, blonde long hair, smiling, looking at camera, indoor, wooden table, bookshelf in background, soft lighting, studio style.
person_0007
Prompt: person_0007, person_0007_vulva, person_0007_anus, nude, adult, light build, front view, close-up, legs spread wide, feet on table, hand on chest, beaded bracelet on left wrist, blonde long hair, smiling, looking at camera, indoor, wooden table, bookshelf in background, soft lighting, studio style.
The variants are not interchangeable. Use the 2512 adapter only with the 2512 base and the 2.1 adapter only with the 2.1 base. The scripts are downloadable references, not a promise that every Diffusers/Viggle release or GPU will behave identically.
| Base and workflow | Reference |
|---|---|
| Qwen Image 2512, higher VRAM | reference_2512.py |
| Qwen Image 2512, sequential CPU offload | reference_2512_low_vram.py |
| Qwen Image 2.1 + Viggle, higher VRAM | reference_21_viggle.py |
| Qwen Image 2.1 + Viggle, experimental low-VRAM | reference_21_viggle_low_vram.py |
The low-VRAM 2.1 variant still requires a CUDA GPU for inference. Sequential CPU offload keeps inactive model components in system RAM and moves them to the GPU as needed; it does not run generation entirely on the CPU. It also uses smaller image dimensions. This variant is untested with the final weight, may require considerable system RAM and current compatible library versions, and will be slower than the higher-VRAM reference.
This model focuses on adult anatomy and regional visual realism; it is not a guarantee of correct anatomy or identity in every image. Difficult viewpoints, occlusion, extreme poses, and close crops remain challenging. Some sample prompts intentionally push those limits. Check outputs rather than trusting tokens alone. The 2.1 branch may receive further training if testing supports it.
For clearly adult, lawful, consensual image generation and research only. Do not use it for minors, non-consensual content, sexualized depictions of identifiable real people without permission, harassment, or deception.
Non-commercial use only. Commercial use or redistribution requires prior permission from the author.