This repository contains a collection of FP16 precision LoRA (Low-Rank Adaptation) adapters for Stable Diffusion XL (SDXL) models, focused on NSFW (Not Safe For Work) content generation.

Model Description

LoRA adapters provide efficient fine-tuning of SDXL models by training only a small subset of parameters, enabling:

Style Transfer: Apply specific artistic styles or aesthetic preferences
Character Consistency: Maintain consistent character appearances across generations
Concept Learning: Train on specific concepts, objects, or themes
Memory Efficiency: Small file sizes (typically 10-500MB per LoRA)
Composability: Stack multiple LoRAs for combined effects

Precision: FP16 (Float16) for balance between quality and file size Base Model: Stable Diffusion XL Base 1.0 Content: NSFW-focused LoRA adapters

Repository Contents

sdxl-fp16-loras-nsfw/
├── loras/
│   └── sdxl/          # SDXL LoRA adapters (.safetensors)
└── README.md          # This file

Model Files

Status: Repository is currently empty. Model files will be added in .safetensors format.

Expected structure:

loras/sdxl/*.safetensors - Individual LoRA adapter files
Typical size: 10-500MB per LoRA file
Format: SafeTensors (secure, efficient weight storage)

Hardware Requirements

VRAM Requirements

Minimum: 8GB VRAM (with optimizations)
Recommended: 12GB VRAM (comfortable generation)
Optimal: 16GB+ VRAM (batch processing, multiple LoRAs)

Disk Space

Per LoRA: 10-500MB (typical: 50-150MB)
Recommended: 5GB+ for collection storage

System Requirements

OS: Windows 10/11, Linux, macOS
Python: 3.10+
CUDA: 11.8+ (for NVIDIA GPUs)
RAM: 16GB+ system RAM recommended

Usage Examples

Basic Usage with Diffusers

python

from diffusers import DiffusionPipeline
import torch

# Load base SDXL model
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipe.to("cuda")

# Load LoRA weights
lora_path = "E:/huggingface/sdxl-fp16-loras-nsfw/loras/sdxl/example_lora.safetensors"
pipe.load_lora_weights(lora_path)

# Generate image
prompt = "your prompt here"
image = pipe(
    prompt=prompt,
    num_inference_steps=30,
    guidance_scale=7.5,
    cross_attention_kwargs={"scale": 0.8}  # LoRA strength (0.0-1.0)
).images[0]

image.save("output.png")

Multiple LoRA Loading

python

# Load multiple LoRAs with different strengths
pipe.load_lora_weights(
    "E:/huggingface/sdxl-fp16-loras-nsfw/loras/sdxl/style_lora.safetensors",
    adapter_name="style"
)
pipe.load_lora_weights(
    "E:/huggingface/sdxl-fp16-loras-nsfw/loras/sdxl/character_lora.safetensors",
    adapter_name="character"
)

# Set adapter weights
pipe.set_adapters(["style", "character"], adapter_weights=[0.7, 0.8])

# Generate with combined LoRAs
image = pipe(prompt).images[0]

Memory-Efficient Generation

python

# Enable memory optimizations
pipe.enable_model_cpu_offload()
pipe.enable_vae_slicing()
pipe.enable_attention_slicing()

# Generate with reduced memory footprint
image = pipe(
    prompt,
    num_inference_steps=25,
    guidance_scale=7.5
).images[0]

ComfyUI Integration

Place LoRA files in: ComfyUI/models/loras/
In ComfyUI workflow:
- Add "Load LoRA" node
- Connect to model chain
- Set strength (0.0-1.0)
- Generate images

Automatic1111/Forge Integration

Place LoRA files in: stable-diffusion-webui/models/Lora/
In prompt, use: <lora:filename:strength>
- Example: beautiful portrait <lora:style_lora:0.8>
Adjust strength value (0.1-1.0) for effect intensity

Model Specifications

Architecture Details

Base Architecture: SDXL (Stable Diffusion XL)
Adapter Type: LoRA (Low-Rank Adaptation)
Precision: FP16 (16-bit floating point)
Format: SafeTensors
Rank: Typically 8-128 (varies by LoRA)
Alpha: Model-specific (check individual LoRA metadata)

Training Details

Base Model: Stable Diffusion XL Base 1.0
Resolution: 1024x1024 (SDXL native)
Content Type: NSFW-focused training data
Optimization: LoRA efficient fine-tuning

Supported Resolutions

Native: 1024x1024
Supported: 512x512 to 2048x2048
Aspect Ratios: 1:1, 16:9, 9:16, 4:3, 3:4, and custom

Performance Tips

Optimization Strategies

LoRA Strength: Start with 0.6-0.8, adjust based on results
Inference Steps: 25-40 steps for quality (lower = faster)
Guidance Scale: 7-9 for balanced results
VAE Slicing: Enable for memory efficiency
CPU Offload: Use for <12GB VRAM systems

Quality Improvements

Use multiple LoRAs strategically (style + concept)
Adjust strengths independently for fine control
Combine with textual inversion embeddings
Use high-quality prompts with detail keywords
Enable xformers for faster generation (if available)

Memory Management

python

# Aggressive memory optimization
pipe.enable_model_cpu_offload()
pipe.enable_vae_tiling()
pipe.enable_attention_slicing(slice_size=1)

# Clear cache between generations
import gc
gc.collect()
torch.cuda.empty_cache()

License

This repository follows the OpenRAIL++ License, which permits:

Commercial Use: Yes, with responsibility requirements
Redistribution: Yes, under same license terms
Modification: Yes, derivative works allowed
Attribution: Recommended but not required

Important Restrictions:

May not be used to generate illegal content
May not be used to harm, exploit, or deceive individuals
Users are responsible for downstream applications
Content warnings required for NSFW outputs

Full license: CreativeML Open RAIL++-M License

Content Warning

This repository contains adapters designed for NSFW (Not Safe For Work) content generation. Users must:

Be 18+ years of age in their jurisdiction
Comply with local laws regarding adult content
Use responsibly and ethically
Implement appropriate content filters in production systems
Not generate illegal or harmful content

Citation

If you use these models in research or production, please cite:

bibtex

@misc{sdxl-fp16-loras-nsfw,
  title={SDXL FP16 LoRA Collection - NSFW},
  author={Community Contributors},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/sdxl-fp16-loras-nsfw}}
}

@article{rombach2022stable,
  title={High-Resolution Image Synthesis with Latent Diffusion Models},
  author={Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj{\"o}rn},
  journal={CVPR},
  year={2022}
}

@article{hu2021lora,
  title={LoRA: Low-Rank Adaptation of Large Language Models},
  author={Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu},
  journal={arXiv preprint arXiv:2106.09685},
  year={2021}
}

Related Resources

Official Documentation

Community Resources

Tools and Interfaces

ComfyUI - Node-based interface
Automatic1111 WebUI - Popular web interface
SD.Next - Advanced fork with features
InvokeAI - Professional interface

Support and Contact

For issues, questions, or contributions:

Issues: Report on GitHub repository issues page
Community: Hugging Face model discussions
Updates: Watch repository for new LoRA additions

Repository Status: Empty - Awaiting model file additions

Last Updated: 2025-10-14

README Version: v1.2

wangkanai/sdxl-fp16-loras-nsfw

SDXL FP16 LoRA Collection - NSFW

Model Description

Repository Contents

Model Files

Hardware Requirements

VRAM Requirements

Disk Space

System Requirements

Usage Examples

Basic Usage with Diffusers

Multiple LoRA Loading

Memory-Efficient Generation

ComfyUI Integration

Automatic1111/Forge Integration

Model Specifications

Architecture Details

Training Details

Supported Resolutions

Performance Tips

Optimization Strategies

Quality Improvements

Memory Management

License

Content Warning

Citation

Related Resources

Official Documentation

Community Resources

Tools and Interfaces

Support and Contact

wangkanai/
sdxl-fp16-loras-nsfw