Chinese-Female-NSFW-LoRA-v1

⛔ DEPRECATED — THIS PREVIEW PROJECT HAS BEEN DISCONTINUED

This repository is no longer maintained and should not be used as the current release.
Continue to the replacement project: Chinese Female NSFW X LoRA v1

About the uploaded legacy checkpoints

Several surviving checkpoints were added to this archived repository after a user asked whether the old weights were still available:

• zl_cn_female_x_v1_000012500.safetensors — 12,500 steps
• zl_cn_female_x_v1_000015000.safetensors — 15,000 steps
• zl_cn_female_x_v1_000017500.safetensors — 17,500 steps

These files are provided as unsupported historical checkpoints, not as a newly selected final release. I no longer have the original per-image generation records, so I cannot reliably determine which exact checkpoint produced each example shown on the archived page. The examples have survived; their paperwork has not.

Please do not assume that all examples came from the highest-step file—or from the same file at all. If you want to use these checkpoints, compare them directly with the same prompts, seeds and settings, then choose the one that best suits your workflow. For the maintained model and current documentation, use the replacement project linked above.

---

<details> <summary><strong>View the archived preview page</strong></summary>

Chinese Female NSFW LoRA v1 (Coming Soon)

---

---

🚀 It's time to show the real power of AI

For years, AI-generated images have often been recognizable by their artificial appearance:

• overly perfect faces
• synthetic skin
• unrealistic proportions
• unnatural lighting
• the unmistakable "AI look"

This project explores a different direction:

Teaching AI to understand reality.

From natural skin texture to realistic proportions, from imperfect lighting to subtle photographic details — the goal is to reduce the gap between AI-generated images and real photography.

When AI truly learns the patterns of reality, the boundary between generated images and photographs becomes increasingly difficult to distinguish.

---

🎉 Public Release

The model is entering its final release stage.

After multiple rounds of training, evaluation, and refinement, the final version will be publicly available soon.

The release version focuses on:

• photorealistic image generation
• realistic body characteristics
• natural photographic details
• improved consistency
• reduced AI artifacts

---

Overview

Chinese Female NSFW LoRA v1 is a photorealistic LoRA designed for Qwen Text to Image.

The primary goal of this project is not simply generating images.

It is exploring how AI models can learn the subtle visual patterns that make real photographs look real.

This LoRA focuses on:

• natural skin texture
• realistic body proportions
• authentic visual imperfections
• natural lighting response
• photographic depth and atmosphere

The objective is simple:

Make AI images look less like AI, and closer to reality.

---

Release Notes

v1.0

Initial public release.

Highlights:

• Trained on 10,000+ real-world photographs
• Optimized for Qwen Text-to-Image workflows
• Improved photorealistic appearance
• Better preservation of natural proportions
• Enhanced skin texture and lighting consistency
• Reduced synthetic AI appearance
• Improved prompt understanding and editing consistency

---

Final Training Configuration

---

Training History

The model was trained on:

NVIDIA RTX PRO 6000 Blackwell 96GB GPU

The initial training target was 25,000 steps.

After evaluating intermediate checkpoints, the training process was extended to 100,000 steps to improve:

• realism
• anatomical consistency
• photographic details
• image quality
• prompt response

The final checkpoint was selected after extensive testing and comparison.

---

Reality Comparison

The following comparison demonstrates the evolution of AI image realism.

The same concept was generated using different Qwen generations and acceleration methods.

These examples show how:

• base model improvements
• inference optimization
• specialized LoRA training

can progressively reduce the traditional AI-generated appearance.

---

Qwen Text-to-Image 2508 (Original) (Weight 1, 20 Steps, cfg 3)

---

Qwen Text-to-Image 2512 (Weight 0.6, 20 Steps, cfg 3)

---

Qwen Text-to-Image 2508 (Original) + Lightning (Weight 1, 4 Steps, cfg 1)

---

Qwen Text-to-Image 2512 + Lightning (Weight 0.6, 4 Steps, cfg 1)

---

Qwen Text-to-Image 2508 (Original) (Weight 1, 20 Steps, cfg 3)

---

Qwen Text-to-Image 2512 (Weight 0.6, 20 Steps, cfg 3)

---

Qwen Text-to-Image 2508 (Original) + Lightning (Weight 1, 4 Steps, cfg 1)

---

Qwen Text-to-Image 2512 + Lightning (Weight 0.6, 8 Steps, cfg 1)

---

Personal Recommendation

There is no absolute winner among these four configurations.

The best choice depends on your own preference between:

• image quality
• generation speed
• model characteristics
• workflow requirements

For my personal preference, I recommend:

Qwen Text-to-Image 2508 (original) + Lightning (4/8 Steps)

This combination provides a very good balance between:

• realistic appearance
• natural details
• generation speed
• overall visual consistency

Although newer base models provide improvements in certain areas, the 2508 + Lightning workflow can still produce excellent results with significantly faster inference.

The best configuration is ultimately a matter of personal preference and creative workflow.

---

Recommended Environment

This LoRA is designed primarily for:

• Qwen Text-to-Image (original)
• Diffusers workflow
• High-quality text-to-image generation

Recommended:

• NVIDIA GPU with sufficient VRAM
• FP16 inference
• Qwen Image compatible pipeline

---

Examples

Example 1

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, A young woman with long black hair is lying on her back in the trunk of a car. She is wearing a white blouse and white tights. Her legs are spread apart and her underwear has been pulled down, exposing her genital area. The trunk is lined with cardboard and there is a piece of cloth on the cardboard. The car's interior is black and the trunk is open. The woman appears to be of East Asian descent. The lighting in the image is dim, suggesting it is nighttime.

</details>

---

Example 2

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, A young Asian woman with fair skin and black hair pulled back into a messy bun is posing for the camera. She is wearing round, black-rimmed glasses and a white tank top that is pulled down to expose her bare breasts. Her arms are raised above her head, and she is smiling. She is sitting on a bed with a light green sheet. The background is a plain white wall.

</details>

---

Example 3

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, A woman is taking a mirror selfie in a gym. She is wearing a light purple sports bra. The sports bra has white straps and the leggings have a hole to reveal her pussy,  She is holding a smartphone in her right hand to take the picture. In the background, there are various gym equipment including red and white barbells, a green exercise mat, and a weight scale. There is also a person in the background, but their face is not. The gym appears to be well-lit and clean.

</details>

---

Example 4

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, The image is a close-up photograph focusing on the buttocks and genital area of a nude individual. The person is lying on their side on a bed with white sheets. The background features a wooden headboard with vertical slats, suggesting a bedroom setting. The skin tone is light, and the individual's body is smooth with minimal hair. The lighting is soft, highlighting the natural contours of the body. The photograph captures a private and intimate moment, emphasizing the natural human form.

</details>

---

Example 5

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, The image is a mirror selfie taken in a modern bathroom. The main focus is on a young man and woman standing in front of a large, round mirror with a light gray frame. The man, who has short black hair and is shirtless, stands behind the woman, his arms wrapped around her waist. The woman, who has long, dark brown hair, is holding a red phone in her right hand, taking the selfie. She is wearing a white bikini bottom and is topless, covering her chest with her left hand. The bathroom has a sleek design with white walls and a glass shower door in the background. The mirror reflects the couple's image clearly, showing the man's face and the woman's profile. The overall setting is bright and clean, with a neutral color palette dominated by whites and grays.

</details>

---

Example 6

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, This is a photograph of a woman in a parking garage. She is positioned on all fours, facing away from the camera, with her buttocks prominently displayed. The woman has long, straight black hair and is wearing a short black dress that is lifted to reveal her bare buttocks. She is also wearing black stockings. The parking garage is empty, with concrete floors and white and yellow parking lines. The walls are painted white and orange, and there are several pillars with the letter "C" on them. The lighting is bright and even, suggesting it is daytime. The image is taken from a low angle, emphasizing the woman's pose and the details of her body.

</details>

---

Example 7

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, A young woman is sitting on a yellow bicycle, which appears to be a shared bike, as indicated by the text "美团App扫码" (Meituan app scan code) on its frame. She is wearing a light pink, short-sleeved blouse with a large bow at the collar and nude for lower body. Her legs are bare, and the pussy is exposed, and she is wearing black shoes with a strap. She has a white shoulder bag slung over her right shoulder. The scene is set at night, with streetlights illuminating the area, and trees are in the background. The woman is holding the handlebars of the bicycle, and her left hand is resting on the seat.

</details>

---

Example 8

<p align="center">

Prompt

<details> <summary>Show Prompt</summary>

zl_cn_female_x_v1, The image features a nude woman with long, dark hair tied back in a ponytail, kneeling on a rooftop. She is positioned with her back to the camera, her arms raised above her head, and her legs slightly apart. The woman is wearing light-colored, high-heeled sandals. Her skin tone is light, and her body is slender with a buttock and back. The rooftop is made of dark, weathered material, and there is a large, cylindrical, blue and white object, possibly a ventilation duct or a large pipe, positioned to the left of the woman. The background includes a cloudy sky and a distant cityscape with buildings in the far background. The overall setting appears to be an industrial or utilitarian rooftop

</details>

---

🎁 A Small Gift for Creators

AI generation is not a button that produces the same result every time.

The same model, the same prompt, and the same settings can create completely different results depending on:

• imagination
• prompt design
• personal taste
• creative direction

To help creators explore this LoRA, a small collection of prompts used during training and testing is provided.

You can download them here:

📄 Training Prompt Collection

Download prompts

These prompts are not "the correct way" to use this model.

They are simply examples of how the model learned visual concepts.

Feel free to modify them, combine them, break them, or completely ignore them.

Sometimes the most interesting results come from the prompts nobody expected.

AI provides the canvas.
Your imagination decides what appears on it.

Have fun creating.

---

Limitations

Like all generative AI models, this LoRA has limitations.

Because the training dataset contains real-world photographs from diverse sources, some details may be affected by:

• limited representation
• occlusion
• motion blur
• unusual camera angles
• extreme poses

The model may occasionally generate local inaccuracies.

Future versions may further improve these areas through additional data and training.

---

Content Notice

This model was trained on a dataset containing real-world photographs, including NSFW content.

The model may generate content that is not suitable for all audiences.

Please use this model responsibly and comply with:

• applicable laws
• platform policies
• community guidelines

---

Notice

The purpose of this project is to explore AI image generation technology and improve visual realism.

This model is intended for:

• research
• experimentation
• creative image generation

Users are solely responsible for their generated content.

---

License

This project is released under:

Chinese Female Body LoRA License v1.0

Please read the full license before downloading or using this model:

License: LICENSE.md

By downloading or using this model, you agree:

• This model contains NSFW knowledge learned from the training dataset.
• Commercial use is strictly prohibited.
• You must comply with applicable laws and regulations.
• You are solely responsible for generated content.

Restrictions

• ❌ Commercial use is strictly prohibited.
• ❌ Do not use this model for unlawful, harmful, or abusive purposes.
• ❌ Do not use this model to violate privacy or individual rights.

---

Compliance

Users are solely responsible for ensuring that their use of this model and any generated content complies with:

• laws and regulations of their country or region
• platform terms and policies
• copyright and privacy requirements

---

Acknowledgements

Special thanks to the Qwen Image team, the open-source AI community, ChatGPT, and Codex for making this project possible.

AI development is rarely a straight line.

During this project, there were many experiments, wrong turns, unexpected results, and moments where things did not work as planned.

ChatGPT and Codex were not always perfect — sometimes they pointed me in the wrong direction, and sometimes they helped me discover a better one through trial and error.

But they were always valuable companions in the process:

• helping explore ideas
• debugging problems
• explaining concepts
• suggesting new directions

The most important part is not getting every answer right immediately.

It is having the ability to keep experimenting, learning, and improving.

Without modern AI assistants and open-source tools, this project would have taken much longer to complete.

The future of AI image generation is not only about creating new images.

It is about learning reality itself.

---

Python Inference Reference

Install or update the required packages:

pip install -U diffusers transformers accelerate safetensors peft

The following example loads the original Qwen Text-to-Image base model, the official 8-step Qwen-Image-Lightning LoRA, and this LoRA together.

The final LoRA weight filename has intentionally been left blank. Set LORA_WEIGHT after the model file is uploaded.

import math
import torch
from diffusers import FlowMatchEulerDiscreteScheduler, QwenImagePipeline


BASE_MODEL = "Qwen/Qwen-Image"

LORA_REPO = "perpetual3x/Chinese-Female-NSFW-LoRA-v1"
LORA_WEIGHT = ""

LIGHTNING_REPO = "lightx2v/Qwen-Image-Lightning"
LIGHTNING_WEIGHT = "Qwen-Image-Lightning-8steps-V2.0-bf16.safetensors"

STYLE_ADAPTER = "Chinese-Female-NSFW-LoRA"
LIGHTNING_ADAPTER = "Qwen-Image-Lightning-8steps"


# Qwen-Image-Lightning uses a distilled FlowMatch scheduler configuration.
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 = QwenImagePipeline.from_pretrained(
    BASE_MODEL,
    scheduler=scheduler,
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

# Load the 8-step Lightning acceleration LoRA.
pipe.load_lora_weights(
    LIGHTNING_REPO,
    weight_name=LIGHTNING_WEIGHT,
    adapter_name=LIGHTNING_ADAPTER,
)

if not LORA_WEIGHT:
    raise ValueError("Set LORA_WEIGHT to the uploaded .safetensors filename before running inference.")

# Load Chinese Female NSFW LoRA.
pipe.load_lora_weights(
    LORA_REPO,
    weight_name=LORA_WEIGHT,
    adapter_name=STYLE_ADAPTER,
)

# Enable both adapters. Adjust the style weight according to preference.
pipe.set_adapters(
    [LIGHTNING_ADAPTER, STYLE_ADAPTER],
    adapter_weights=[1.0, 1.0],
)

prompt = (
    "zl_cn_female_x_v1, a photorealistic portrait of an adult Chinese woman, "
    "natural skin texture, realistic proportions, soft photographic lighting, "
    "authentic details, high-resolution photography"
)

negative_prompt = (
    "low resolution, low quality, deformed limbs, malformed fingers, "
    "oversaturated colors, waxy appearance, missing facial details, "
    "overly smooth skin, obvious AI-generated appearance, chaotic composition, "
    "blurry text, distorted text"
)

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

with torch.inference_mode():
    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-LoRA-v1.png")
print("Saved: Chinese-Female-NSFW-LoRA-v1.png")

Keep the 8-step Lightning adapter enabled when using this configuration. The style adapter weight is a starting point, not a fixed rule; adjust it according to your preferred visual strength.

Lightning reference: lightx2v/Qwen-Image-Lightning

Official base-model reference: Qwen/Qwen-Image

---

Related Projects

A LoRA project exploring Japanese Female Aesthetics:

Japanese Female Aesthetics NSFW LoRA v1

</details>