Moss Image NSFW Uncensored LoRA

An adult photographic-content LoRA for MOSS-Image-SFT or MOSS-Image-Base, with MOSS-Image-SFT preferred for inference. It is tuned for explicit adult portraiture, nude poses, indoor scenes, and close photographic compositions.

The repository-root adapter is now the focused 1,750-step checkpoint trained on 130 images; start at scale 0.8. The former 2,000-step, 221-image adapter is preserved under original-checkpoint-2000/ and remains available at its recommended scale of 0.7.

A GGUF conversion (44% less VRAM on Q8 and 66% less VRAM on Q4) is also available here: https://huggingface.co/axiomofmind/Moss-image-nsfw-uncensored-lora-GGUF which can be used with MOSS-Image-SFT-GGUF

Adult-content notice: This repository contains explicit sexual imagery and prompts depicting adults. It is intended only for adults in jurisdictions where such content is lawful. Do not use it to depict minors or age-ambiguous subjects, sexual abuse, non-consensual acts, exploitation, or sexualized representations of real people without their explicit consent.

New merged model with alternative checkpoint (scale 0.8)

For convenience, merged-model/ contains a complete standalone Diffusers model with the now-default focused 1,750-step checkpoint fused into OpenMOSS-Team/MOSS-Image-SFT at scale 0.8. This checkpoint was trained on the focused 130-image dataset. The bundle replaces the previous original-checkpoint scale-0.7 merge, so the repository carries only one complete merged model. It was reloaded and validated at 1024 x 1024 without a separate adapter. The merged model has a fixed strength and is about 25.4 GiB; users who want a smaller download or adjustable strength should use the LoRA files.

The merged model still expects the moss_xvstyle trigger phrase at the beginning of the prompt. Download only the merged bundle with:

hf download axiomofmind/Moss-image-nsfw-uncensored-lora \
  --include "merged-model/**" \
  --local-dir ./Moss-image-nsfw-uncensored-lora

Then load ./Moss-image-nsfw-uncensored-lora/merged-model as the local model path with the MOSS custom Diffusers pipeline. No LoRA path or runtime LoRA scale is needed.

Merge at a different strength

The included scripts/merge_moss_lora.py utility can create another standalone model from the original SFT model and the default focused checkpoint at the repository root. For example, to bake it at scale 0.9 instead of the bundled 0.8:

hf download OpenMOSS-Team/MOSS-Image-SFT --local-dir ./MOSS-Image-SFT
hf download axiomofmind/Moss-image-nsfw-uncensored-lora \
  adapter_config.json \
  pytorch_lora_weights.safetensors \
  scripts/merge_moss_lora.py \
  --local-dir ./Moss-image-nsfw-uncensored-lora

python ./Moss-image-nsfw-uncensored-lora/scripts/merge_moss_lora.py \
  --model-path ./MOSS-Image-SFT \
  --lora-path ./Moss-image-nsfw-uncensored-lora \
  --output ./Moss-image-nsfw-uncensored-focused-1750-merged-s0p9 \
  --scale 0.9 \
  --device cuda:0 \
  --preferred-trigger moss_xvstyle

The focused checkpoint was evaluated at scales 0.7, 0.8, and 0.9; substitute any of those values and use a matching output directory name. After merging, the chosen scale is permanent. Continue using moss_xvstyle in prompts.

Merge the original (primary) checkpoint

The original 2,000-step checkpoint trained on 221 images is preserved under original-checkpoint-2000/. Download it with the merge script:

hf download axiomofmind/Moss-image-nsfw-uncensored-lora \
  original-checkpoint-2000/adapter_config.json \
  original-checkpoint-2000/pytorch_lora_weights.safetensors \
  scripts/merge_moss_lora.py \
  --local-dir ./Moss-image-nsfw-uncensored-lora

Point --lora-path at the original checkpoint directory. This example bakes it at its recommended scale of 0.7:

python ./Moss-image-nsfw-uncensored-lora/scripts/merge_moss_lora.py \
  --model-path ./MOSS-Image-SFT \
  --lora-path ./Moss-image-nsfw-uncensored-lora/original-checkpoint-2000 \
  --output ./Moss-image-nsfw-uncensored-primary-merged-s0p7 \
  --scale 0.7 \
  --device cuda:0 \
  --preferred-trigger moss_xvstyle

The original checkpoint remains available as adjustable PEFT and ComfyUI LoRA files; it is not included as a second full merged model.

ComfyUI compatibility

The root pytorch_lora_weights.safetensors uses Diffusers / PEFT keys such as transformer.layers.0.proj.q.lora_A.weight. ComfyUI's native MOSS-Image loader addresses the same modules under diffusion_model.layers..., so loading the root file directly can report lora key not loaded.

Use one of the native ComfyUI variants instead:

1. Update ComfyUI to a version with native MOSS-Image support; commit b1a247c or newer contains the required loader code.
2. Use the default focused file at model strength 0.7, 0.8, or 0.9; 0.8 is the suggested starting point. The archived original checkpoint is recommended at 0.7.
3. Put the selected .safetensors file in ComfyUI/models/loras/.
4. Use ComfyUI's standard LoRA loader with clip strength 0.0.
5. Begin the prompt with moss_xvstyle; local comparisons were slightly better with the trigger than without it.

Download the default focused ComfyUI variant with:

hf download axiomofmind/Moss-image-nsfw-uncensored-lora \
  MOSS-Image-NSFW-Uncensored-LoRA-ComfyUI.safetensors \
  --local-dir ./ComfyUI/models/loras

Download the original 2,000-step ComfyUI variant with:

hf download axiomofmind/Moss-image-nsfw-uncensored-lora \
  original-checkpoint-2000/MOSS-Image-NSFW-Uncensored-LoRA-ComfyUI.safetensors \
  --local-dir ./Moss-image-nsfw-uncensored-lora

Then copy the downloaded original-checkpoint .safetensors file into ComfyUI/models/loras/.

The conversion is namespace-only: transformer. becomes diffusion_model.. Tensor values, shapes, dtypes, rank, and alpha behavior are unchanged. All 504 converted tensors were checked against the 252 corresponding MOSS-Image-SFT modules. The reproducible converter is scripts/convert_moss_lora_for_comfyui.py.

The separate GGUF adapter is for stable-diffusion.cpp and is not this native ComfyUI format. The bundled merged-model/ avoids runtime LoRA key matching when used as a Diffusers model, but the sharded Diffusers directory is not a direct replacement for the native ComfyUI LoRA file in a standard ComfyUI model-loading workflow.

Recommended settings

• Trigger: moss_xvstyle
• Preferred underlying model: OpenMOSS-Team/MOSS-Image-SFT
• Also compatible with: OpenMOSS-Team/MOSS-Image-Base
• Default checkpoint: 1,750 steps from the focused 2,000-step rerun
• Training images: 130
• LoRA scale: 0.7, 0.8, or 0.9; start with 0.8
• Inference steps: 30
• Guidance scale: 3.0
• Guidance testing: only 3.0 was evaluated; other guidance values were not tested
• Resolution: 1024 x 1024
• Negative prompt: empty

Place moss_xvstyle at the beginning of the prompt. The adapter was trained against MOSS-Image-Base, which remains the training parent recorded in adapter_config.json. Its transformer targets are compatible with MOSS-Image-SFT, and the focused checkpoint is now published under the canonical root filenames.

Original checkpoint (2,000 steps)

The former repository-root checkpoint is preserved under original-checkpoint-2000/. It was trained on 221 images and selected at step 2,000. Use it when you prefer the original behavior shown in the comparison tables.

Recommended settings for the original checkpoint:

• Weights: original-checkpoint-2000/pytorch_lora_weights.safetensors
• ComfyUI weights: original-checkpoint-2000/MOSS-Image-NSFW-Uncensored-LoRA-ComfyUI.safetensors
• Trigger: moss_xvstyle
• Preferred underlying model: OpenMOSS-Team/MOSS-Image-SFT
• LoRA scale: 0.7
• Inference steps: 30
• Guidance scale: 3.0
• Resolution: 1024 x 1024
• Negative prompt: empty

Download only the original checkpoint with:

hf download axiomofmind/Moss-image-nsfw-uncensored-lora \
  --include "original-checkpoint-2000/**" \
  --local-dir ./Moss-image-nsfw-uncensored-lora

The Diffusers example below uses the default focused checkpoint at the repository root. To use the original checkpoint, change config_path and weights_path to the corresponding files inside original-checkpoint-2000/, and set lora_scale to 0.7.

Examples

These examples mirror the selected model-card widget images. All use MOSS-Image-SFT, image seed 42, 30 inference steps, guidance scale 3.0, and 1024 x 1024 resolution. Eight use the default focused 1,750-step checkpoint at the scale shown; Sample 4 uses the original 2,000-step checkpoint at scale 0.7.

Sample 0 — rear close-up — default focused checkpoint, scale 0.9

Prompt: moss_xvstyle, a close-up photograph of an adult woman with fair skin and long blonde hair, viewed from behind as she bends forward over a gray surface. Her body is positioned with her buttocks prominently displayed, revealing her vulva and labia in a natural, unposed state. She wears a white thong that is pulled down to expose her intimate area. The lighting is soft and natural, highlighting the smooth texture of her skin. The background is blurred, suggesting an indoor setting with neutral tones.

Sample 1 — bedroom mirror scene — default focused checkpoint, scale 0.8

Prompt: moss_xvstyle, A full-body photograph shows an adult woman with a curvaceous body and long brown hair styled in a braid, kneeling on a wooden floor with her back to the camera and her head resting on her arms. Her skin tone is light tan, and she is completely nude, with her large buttocks prominently displayed and her vulva visible between her thighs. She wears a gold anklet and a green floral hair accessory. Her gaze is directed downward, and her expression is neutral. The setting is a bedroom with a large mirror reflecting her pose, a white upholstered chair, and light-colored curtains. The lighting is soft and indoor, and the camera angle is a medium shot from behind.

Sample 2 — crouching room portrait — default focused checkpoint, scale 0.9

Prompt: moss_xvstyle, a nude adult woman with fair skin and shoulder-length light brown hair is photographed in a medium close-up from a slightly low angle. She is crouching with her legs spread wide, exposing her vulva and labia, while her left hand rests on a wooden surface behind her. Her right arm is raised above her head. She has a belly button piercing with a silver stud and is wearing white ankle socks with yellow smiley faces. Her facial expression is neutral, and she gazes directly at the camera. The setting is an indoor room with a white wall, a dark television screen, and a white lamp in the background. The lighting is soft and even, highlighting her smooth skin and large breasts. The image is a photograph.

Sample 3 — reclining POV — default focused checkpoint, scale 0.8

Prompt: moss_xvstyle, an adult woman with long blonde hair styled in loose waves and fair skin is positioned in a close-up POV shot, lying on her back with her legs spread wide to expose her vulva and labia. She wears a black halter-style top that leaves her breasts bare and has a tattoo on her lower abdomen. Her right hand rests between her thighs with fingers near her clitoris, and she wears a ring on her finger. Her mouth is slightly open and she gazes directly at the camera with an aroused expression. The setting appears to be a modern indoor room with light-colored walls, a metal railing, and wooden cabinetry in the background, illuminated by soft overhead lighting. The image is a photograph.

Sample 4 — side-lying close-up — original checkpoint, scale 0.7

Prompt: moss_xvstyle, A close-up photograph of an adult woman lying on her side on a gray fabric surface, her left hand resting on her upper thigh with pink-painted fingernails visible. She has light brown skin and is nude, with her legs bent and pressed together, exposing her vulva and inner thighs. Her gaze is directed toward the camera with a neutral expression. The background is softly blurred, suggesting an indoor setting with natural lighting. The image is a photograph.

Sample 5 — rear bedside pose — default focused checkpoint, scale 0.8

Prompt: moss_xvstyle, A realistic adult photograph of a curvy woman with long dark wavy hair kneeling near the edge of an unmade bed in a rear three-quarter pose. She wears a burgundy lace camisole lifted above her waist and matching panties pulled down around her upper thighs, leaving her buttocks, vulva, and labia visibly exposed. She looks back over her shoulder toward the camera with a calm expression. Soft bedside lamplight mixes with cool evening window light, creating detailed natural skin texture and gentle shadows. The framing is a medium full-body portrait in a tasteful contemporary bedroom.

Sample 6 — reclining sofa side pose — default focused checkpoint, scale 0.8

Prompt: moss_xvstyle, A close natural-light photograph of an adult woman with warm brown skin and short curly black hair reclining on her side across a deep green velvet sofa. She is nude with her upper leg raised and bent, naturally revealing her breasts, inner thighs, vulva, and labia while her lower hand rests loosely on her hip. She looks directly at the camera with a neutral, intimate expression. Diffuse morning light from a nearby window creates soft dimensional shadows, realistic pores, and detailed fabric texture. The composition is a horizontal-feeling three-quarter body portrait captured from sofa height in a quiet living room.

Sample 7 — marble vanity — default focused checkpoint, scale 0.8

Prompt: moss_xvstyle, A realistic frontal photograph of an adult East Asian woman with long straight black hair seated on the edge of a white marble bathroom vanity. She wears an open cream satin robe around her shoulders and is otherwise nude, sitting with her legs apart so her breasts, vulva, and labia are naturally and clearly visible. One hand rests beside her on the counter while the other touches the robe collar. Her expression is calm and self-assured as she looks toward the camera. Soft morning window light and warm vanity bulbs create balanced highlights, realistic skin texture, and gentle shadows in a refined modern bathroom.

Sample 8 — sunlit bed rear pose — default focused checkpoint, scale 0.8

Prompt: moss_xvstyle, A detailed adult photograph of a freckled woman with long copper-red hair posed on her hands and knees across rumpled white bedsheets, viewed from a rear three-quarter angle. She wears a sheer pale-blue lingerie top and matching underwear pulled down below her buttocks, clearly revealing her vulva and labia between her thighs. She turns her head toward the camera with a relaxed expression. Strong but diffused midday sunlight passes through gauzy curtains, producing natural skin tones, visible freckles, soft shadows, and crisp bedding texture in an airy bedroom.

Stock, original, and default checkpoint comparison

The table compares three generation conditions using the same descriptive prompt, image seed 42, 1024 x 1024 resolution, 30 inference steps, and guidance scale 3.0:

• Stock MOSS-Image-SFT: SFT without the LoRA and with only the trigger phrase removed.
• Original checkpoint: MOSS-Image-SFT with the archived 2,000-step adapter trained on 221 images, at scale 0.7 with the trigger included.
• Default focused checkpoint: MOSS-Image-SFT with the repository-root 1,750-step adapter trained on 130 images. The selected examples use scale 0.8 or 0.9, shown in each cell, with the trigger included.

Original and default checkpoint comparison

This table uses matched prompts and image seed 42 at 1024 x 1024, 30 inference steps, and guidance scale 3.0. Stock MOSS-Image-SFT uses no LoRA and omits the trigger. Both LoRA columns use moss_xvstyle:

• Original checkpoint: the archived 2,000-step adapter trained on 221 images, used at scale 0.7.
• Default focused checkpoint: the repository-root 1,750-step adapter trained on 130 images, shown at scale 0.8.

Use with Diffusers and PEFT

MOSS-Image uses a custom Diffusers pipeline. Attach the adapter to its transformer with PEFT:

import json

import torch
from diffusers import DiffusionPipeline
from huggingface_hub import hf_hub_download
from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
from peft.tuners.lora.layer import LoraLayer
from safetensors.torch import load_file

repo_id = "axiomofmind/Moss-image-nsfw-uncensored-lora"
base_model = "OpenMOSS-Team/MOSS-Image-SFT"  # Preferred; Base is also supported.
lora_scale = 0.8

config_path = hf_hub_download(repo_id, "adapter_config.json")
weights_path = hf_hub_download(repo_id, "pytorch_lora_weights.safetensors")

with open(config_path, encoding="utf-8") as file:
    config = json.load(file)

pipe = DiffusionPipeline.from_pretrained(
    base_model,
    custom_pipeline="pipeline_moss_image",
    trust_remote_code=True,
    torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
)
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)

inject_adapter_in_model(
    LoraConfig(
        r=config["r"],
        lora_alpha=config["lora_alpha"],
        target_modules=config["target_modules"],
    ),
    pipe.transformer,
)

state = {
    key.removeprefix("transformer."): value
    for key, value in load_file(weights_path, device="cpu").items()
}
_, unexpected = set_peft_model_state_dict(pipe.transformer, state)
if unexpected:
    raise RuntimeError(f"Unexpected LoRA tensors: {unexpected[:5]}")

for layer in pipe.transformer.modules():
    if isinstance(layer, LoraLayer):
        layer.set_scale("default", lora_scale)

pipe.transformer.to(torch.bfloat16)
pipe.to("cuda")

prompt = (
    "moss_xvstyle, an explicit photograph of an adult woman in a softly lit "
    "bedroom, natural skin texture, close composition."
)

image = pipe(
    prompt,
    negative_prompt="",
    width=1024,
    height=1024,
    num_inference_steps=30,
    guidance_scale=3.0,
    uncond_skip_layers=[22],
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]

image.save("moss_xvstyle.png")

Checkpoint details

• Training base model: OpenMOSS-Team/MOSS-Image-Base
• Preferred inference model: OpenMOSS-Team/MOSS-Image-SFT
• Default checkpoint: 1,750 steps from a 2,000-step focused rerun
• Training images: 130
• Resolution: 1024
• LoRA rank / alpha: 16 / 16
• Target modules: transformer attention projections and feed-forward projections
• Optimizer: AdamW
• Learning rate: 1e-4
• Scheduler: constant with warmup
• Batch size / gradient accumulation: 1 / 1
• Weight decay: 0.01
• Training seed: 42
• Caption dropout: 0
• Random horizontal flip: disabled
• Recommended LoRA scale: 0.7, 0.8, or 0.9
• Guidance evaluation: 3.0 only; other guidance values were not tested

Original archived checkpoint:

• Path: original-checkpoint-2000/
• Checkpoint: 2,000 steps
• Training images: 221
• Recommended LoRA scale: 0.7

Intended use, restrictions, and limitations

This adapter is intended for lawful creative generation involving clearly adult fictional subjects. Users are responsible for complying with applicable law, platform rules, privacy rights, and consent requirements.

Do not use this adapter for:

• Any depiction of minors or age-ambiguous people in sexual contexts.
• Sexual abuse, exploitation, coercion, trafficking, or non-consensual imagery.
• Sexualized deepfakes or depictions of identifiable real people without explicit consent.
• Harassment, blackmail, doxxing, or attempts to evade legal or platform safeguards.

The model can produce anatomical errors, malformed hands, inconsistent reflections, implausible poses, or outputs that diverge from the prompt. A trigger word does not guarantee a particular style or content level. Review outputs before use or distribution.

License

The adapter files are provided under the Apache License 2.0, matching the base model. This is an independent community LoRA and is not an official OpenMOSS release. The license does not override laws or third-party rights governing generated or source content.