Yes, it is another Tumblr LoRA. Apparently the old JPEGs were not finished haunting my GPU—and I was not finished pretending this counts as archaeology.

Important Realism Limitation: A significant portion of the source dataset consists of selfies and images captured with beauty-camera apps. As a result, the LoRA may sometimes reproduce the less convincing characteristics of that material—such as over-smoothed skin, beauty-filter processing, or phone-camera artifacts—rather than strict photographic realism. This is an important limitation of the model, not an undocumented beauty-filter feature.

TumblrAsian NudeShot NSFW LoRA v1 is an experimental LoRA for Qwen/Qwen-Image-2512 Text-to-Image. It was trained on the Asian subset of a Tumblr-derived adult image collection, with a narrow objective: improve the accuracy and consistency of specific anatomical details.

This is not intended to be a universal realism model. It is a focused training experiment designed to test how much a specialized dataset and detail-aware captions can improve local anatomy while preserving broader photographic generation capability.

Adult-content notice: This repository contains explicit adult material and is intended only for adults. All people described in the examples are adults.

✨ Features

Experimental anatomy-detail enhancement for adult Asian female imagery
Trained on 12,793 images from the Asian portion of a Tumblr-derived dataset
Increased attention to breast and vulva detail in relevant compositions
Separate optional detail tokens for the two anatomy-focused subsets
Supports full-body, medium, portrait, selfie, and close-up compositions
Designed and tested for Qwen-Image-2512 T2I

🧪 Experimental Goal

The base model can already generate the human body, but difficult viewpoints, close framing, occlusion, and small local structures may still produce generic or inaccurate anatomy. This training experiment concentrates additional learning capacity on those details.

The intended improvement is local accuracy, not a guarantee of perfect anatomy. Results still depend heavily on the prompt, seed, composition, resolution, LoRA weight, and base model behavior.

🖥️ Training Details

Hardware

GPU: NVIDIA RTX PRO 6000 Blackwell 96GB

Dataset and Network

Training images: 12,793
Dataset source: Tumblr-derived Asian adult image subset
Batch size: 8
Training steps: 6,000
LoRA rank: 32
Inference base model: Qwen/Qwen-Image-2512 T2I

Trigger Words

Use the main trigger for general generation:

zl_tumblrasia_v1

For vulva-focused detail, use:

zl_tumblrasia_v1, zl_tumblrasiavulva_detail_v1

For breast-focused detail, use:

zl_tumblrasia_v1, zl_tumblrasiabreast_detail_v1

The detail triggers were attached only to matching subsets during training. Use them together with the main trigger, not as replacements for it. They may be unnecessary for distant or fully clothed compositions.

⚠️ Base Model Compatibility

This LoRA is intended for:

Qwen/Qwen-Image-2512 for Text-to-Image generation

Compatibility with Qwen-Image-Edit or unrelated model architectures is not guaranteed. A different base checkpoint may produce weak effects, artifacts, incorrect colors, or a loading error.

🔍 Scope and Limitations

Possible limitations include:

over-smoothed or plastic-looking skin
beauty-filter facial proportions or processing artifacts
inconsistent lighting, color balance, and phone-camera sharpening
compression artifacts and low-resolution Tumblr-era image characteristics
inaccurate hands, fingers, feet, limbs, or body proportions
local anatomy that remains incorrect in difficult angles or heavy occlusion
an occasional conflict between anatomical detail and overall realism

This model improves tendencies; it does not enforce anatomical correctness. Review generated output carefully and regenerate or edit visible defects when necessary.

🖼️ Examples

Six samples are included below. A few arrived with enough anatomical enthusiasm to make an ordinary preview page quietly update its résumé, so those images were converted to black and white.

This is not a secret monochrome feature. It is simply the digital equivalent of dimming the museum lights while continuing to insist that the exhibition is educational. The LoRA remains exactly as shameless; only a few samples learned to behave in grayscale.

All people described in the prompts are adults.

Example 1

Prompt

This is a black and white photograph of an adult Asian woman posing in a provocative manner. She is fully nude with big breasts. Her legs are spread wide apart, and she is bending over, revealing her buttocks. The background is fully dark. Full-body view.

Example 2

Prompt

The image features a nude adult Asian woman sitting on a white cube. She has a slender physique with small breasts and no pubic hair. Her skin tone is light brown, and she has a straight nose and dark eyes. Her hair is long, straight, and dark brown, tied back in a ponytail. She is positioned with her legs spread apart, one foot resting on the floor and the other on the cube. Her right hand rests on her right knee. The background is fully dark, and there is a light blue towel and a pair of silver high-heeled shoes on the floor beside the cube. The lighting is soft, creating a natural and relaxed atmosphere.

Example 3

Prompt

The image is a photograph featuring a nude adult Asian woman standing outdoors against a dark background, possibly a cloudy sky by the seaside. She has long, dark hair that cascades down her back and shoulders. Her body is positioned slightly to the side, with her right hand resting on her hip and her left arm hanging loosely by her side. Her facial expression is one of joy and laughter, with her mouth open and eyes crinkled. She has a medium build with natural breasts and a fully shaved pubic area. The overall tone of the image is soft and natural, with a focus on the woman's relaxed and happy demeanor.

Example 4

Prompt

An adult Asian woman stands nude in front of a gray sky and snow field. She has a medium build and fair skin. Her dark hair is cut in a bob style and she wears a pink beanie hat. She is wearing white knee-high boots with a buckle detail. Her legs are slightly apart. She has a neutral expression on her face.

Example 5

Prompt

This is a black and white photograph of an adult Asian woman. She is wearing a black, sheer fishnet bodysuit that covers her thighs and buttocks, leaving her genitals exposed. The fishnet has a fine mesh pattern and is stretched taut over her skin. Her skin tone is light, and her genitals are visible, including the labia and pubic hair. The background is fully dark. Full-body view.

Example 6

Yes, I used a yoga pose again. I explained the reason before, so regular visitors probably already know. Models change, datasets change, training plans change—but this pose keeps returning like an unofficial unit test with unreasonably good flexibility. If it survives the pose, at least we have learned something.

Prompt

A 20-year-old nude adult Chinese woman with big breasts is performing a challenging yoga pose in an all-black room. She is sitting on the floor with her right leg extended straight up in the air, her left leg bent at the knee, and her left foot flat on the ground. Her right arm is extended upward, holding her right foot, while her left arm rests on the floor for support. The background is dark.

📄 About Training Prompts

This repository does not include a downloadable training prompts.txt. For reusable prompt structures and longer testing examples, refer to the prompts.txt published with the author's other Tumblr LoRA. Adapt those prompts to the subject, composition, and detail trigger required here.

🚀 Reference Code — Qwen-Image-2512 T2I

This example loads Qwen-Image-2512 in BF16, fuses the official community Lightning 4-step adapter, and applies this LoRA at weight 1.0. A high-memory CUDA GPU is recommended.

Installation

pip install -U diffusers transformers accelerate peft safetensors

Python

python

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/TumblrAsian-NudeShot-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "tumblrasia.safetensors"

DEVICE = "cuda"
DTYPE = torch.bfloat16


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="tumblrasia_v1",
)
pipe.set_adapters(["tumblrasia_v1"], adapter_weights=[1.0])

prompt = (
    "zl_tumblrasia_v1, an adult Asian woman, realistic photography, "
    "natural skin texture, full-body composition, soft photographic lighting"
)

negative_prompt = (
    "low quality, blurry, distorted anatomy, malformed hands, extra limbs, "
    "duplicated body parts, plastic skin, CGI, watermark, text"
)

generator = torch.Generator(device=DEVICE).manual_seed(42)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=928,
    height=1328,
    num_inference_steps=4,
    true_cfg_scale=1.0,
    generator=generator,
).images[0]

image.save("tumblrasian_nudeshot_v1.png")

For a detail-focused generation, append one matching optional detail token to the prompt. Start with a LoRA weight between 0.7 and 1.0; excessive weight may amplify beauty-camera artifacts along with the desired detail.

🧩 Low-VRAM Inference (16–24 GB)

Most VRAM is consumed by the Qwen-Image-2512 base model and text encoder, not by this LoRA. Users with lower-memory NVIDIA GPUs can combine 4-bit quantization, CPU offloading, VAE tiling, batch size 1, and a lower output resolution.

Low-VRAM Installation

pip install -U diffusers transformers accelerate peft safetensors bitsandbytes

Low-VRAM Python

python

import torch
from diffusers import DiffusionPipeline
from diffusers.quantizers import PipelineQuantizationConfig


BASE_MODEL_ID = "Qwen/Qwen-Image-2512"
LORA_MODEL_ID = "perpetual3x/TumblrAsian-NudeShot-NSFW-LoRA-v1"
LORA_WEIGHT_NAME = "tumblrasia.safetensors"

quant_config = PipelineQuantizationConfig(
    quant_backend="bitsandbytes_4bit",
    quant_kwargs={
        "load_in_4bit": True,
        "bnb_4bit_quant_type": "nf4",
        "bnb_4bit_compute_dtype": torch.bfloat16,
    },
    components_to_quantize=["transformer", "text_encoder"],
)

pipe = DiffusionPipeline.from_pretrained(
    BASE_MODEL_ID,
    torch_dtype=torch.bfloat16,
    quantization_config=quant_config,
)

pipe.load_lora_weights(
    LORA_MODEL_ID,
    weight_name=LORA_WEIGHT_NAME,
    adapter_name="tumblrasia_v1",
)
pipe.set_adapters(["tumblrasia_v1"], adapter_weights=[0.8])

pipe.enable_model_cpu_offload()
pipe.vae.enable_tiling()

prompt = (
    "zl_tumblrasia_v1, an adult Asian woman, realistic photography, "
    "natural skin texture, soft photographic lighting"
)

negative_prompt = (
    "low quality, blurry, distorted anatomy, malformed hands, extra limbs, "
    "duplicated body parts, watermark, text"
)

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

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=768,
    height=1024,
    num_images_per_prompt=1,
    num_inference_steps=20,
    true_cfg_scale=3.0,
    max_sequence_length=256,
    generator=generator,
).images[0]

image.save("tumblrasian_nudeshot_v1_low_vram.png")

If this still causes an out-of-memory error, replace pipe.enable_model_cpu_offload() with pipe.enable_sequential_cpu_offload(). Sequential CPU offloading uses less GPU memory but is substantially slower.

🔞 Responsible Use

This model is intended exclusively for lawful, consensual, adult creative work. Do not use it to:

create or sexualize minors or youthful-looking persons
create non-consensual or exploitative sexual content
depict an identifiable real person without permission
harass, threaten, exploit, or defame another person
create illegal content or bypass platform rules

Users are responsible for ensuring that their prompts, source material, and generated content comply with applicable laws and platform policies.

License & Copyright Notice

This LoRA may be used for personal and non-commercial purposes only. See the repository LICENSE file for the complete terms.

For commercial licensing inquiries, please contact the author.