wangkanai/sdxl-fp16

Stable Diffusion XL FP16 Model Repository

Local repository containing Stable Diffusion XL (SDXL) checkpoint models in FP16 precision for high-quality text-to-image generation.

Model Description

This repository contains two SDXL checkpoint models optimized for different use cases:

• SDXL Base: Full-featured SDXL 1.0 base model for high-quality image generation with standard inference steps
• SDXL Turbo: Fast inference variant optimized for fewer steps (1-4 steps) while maintaining quality

Both models use FP16 (16-bit floating point) precision, providing a balance between quality and VRAM efficiency.

Repository Contents

E:\huggingface\sdxl-fp16\
├── checkpoints/
│   └── sdxl/
│       ├── sdxl-base.safetensors     (6.94 GB)
│       └── sdxl-turbo.safetensors    (13.88 GB)
├── diffusion_models/
│   └── sdxl/                         (empty - reserved)
└── loras/
    └── sdxl/                         (empty - reserved)

Total Repository Size: ~20.82 GB

Model Files

File	Size	Description
sdxl-base.safetensors	6.94 GB	SDXL 1.0 base checkpoint (FP16)
sdxl-turbo.safetensors	13.88 GB	SDXL Turbo checkpoint (FP16)

Hardware Requirements

SDXL Base

• VRAM: 8GB minimum, 12GB+ recommended
• Disk Space: 7GB for model file
• System RAM: 16GB+ recommended
• GPU: NVIDIA GPU with CUDA support

SDXL Turbo

• VRAM: 12GB minimum, 16GB+ recommended
• Disk Space: 14GB for model file
• System RAM: 16GB+ recommended
• GPU: NVIDIA GPU with CUDA support

Usage Examples

SDXL Base (Standard Quality)

from diffusers import DiffusionPipeline
import torch

# Load SDXL base model from local path
pipe = DiffusionPipeline.from_single_file(
    "E:/huggingface/sdxl-fp16/checkpoints/sdxl/sdxl-base.safetensors",
    torch_dtype=torch.float16
)

pipe.to("cuda")

# Generate image with standard settings
image = pipe(
    prompt="a beautiful mountain landscape at sunset, photorealistic, highly detailed",
    negative_prompt="blurry, low quality, distorted",
    num_inference_steps=50,
    guidance_scale=7.5,
    width=1024,
    height=1024
).images[0]

image.save("output.png")

SDXL Turbo (Fast Generation)

from diffusers import DiffusionPipeline
import torch

# Load SDXL Turbo for fast inference
pipe = DiffusionPipeline.from_single_file(
    "E:/huggingface/sdxl-fp16/checkpoints/sdxl/sdxl-turbo.safetensors",
    torch_dtype=torch.float16
)

pipe.to("cuda")

# Generate with minimal steps (1-4 steps)
image = pipe(
    prompt="a futuristic cityscape at night, neon lights, cyberpunk",
    num_inference_steps=4,  # Turbo optimized for 1-4 steps
    guidance_scale=0.0,     # Turbo works best with guidance_scale=0
    width=1024,
    height=1024
).images[0]

image.save("turbo_output.png")

Memory Optimization

import torch
from diffusers import DiffusionPipeline

# Enable memory-efficient attention
pipe = DiffusionPipeline.from_single_file(
    "E:/huggingface/sdxl-fp16/checkpoints/sdxl/sdxl-base.safetensors",
    torch_dtype=torch.float16
)

# Apply optimizations
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()
pipe.to("cuda")

# Generate with optimized memory usage
image = pipe(
    prompt="your prompt here",
    num_inference_steps=30
).images[0]

Model Specifications

SDXL Base

• Architecture: Latent Diffusion Model with UNet
• Parameters: ~2.6B (UNet backbone)
• Precision: FP16 (16-bit floating point)
• Format: SafeTensors (secure, efficient)
• Resolution: 1024x1024 native, supports 512-2048px
• Text Encoders: Dual CLIP (OpenCLIP ViT-bigG, OpenAI CLIP ViT-L)
• Inference Steps: 30-50 recommended

SDXL Turbo

• Architecture: Adversarial Diffusion Distillation (ADD)
• Parameters: Similar to base with distillation optimizations
• Precision: FP16 (16-bit floating point)
• Format: SafeTensors
• Resolution: 1024x1024 native
• Inference Steps: 1-4 steps (optimized)
• Guidance Scale: 0.0 recommended (classifier-free guidance disabled)

Performance Tips

Speed Optimization

• SDXL Turbo: Use 1-4 steps with guidance_scale=0.0 for fastest generation
• Attention Slicing: Enable with pipe.enable_attention_slicing() for memory efficiency
• VAE Slicing: Enable with pipe.enable_vae_slicing() to reduce VRAM usage
• Lower Resolutions: Use 768x768 or 512x512 for faster generation
• Batch Processing: Process multiple prompts together when VRAM allows

Quality Optimization

• SDXL Base: Use 40-50 steps for highest quality
• Guidance Scale: 7.0-9.0 for base model (higher = more prompt adherence)
• Negative Prompts: Use detailed negative prompts to avoid unwanted elements
• Resolution: 1024x1024 is the native resolution for best results
• Aspect Ratios: Multiples of 64 recommended (1024x768, 768x1024, etc.)

VRAM Management

• 8GB VRAM: Use attention slicing, VAE slicing, lower batch sizes
• 12GB VRAM: Standard settings with optimizations
• 16GB+ VRAM: Can handle higher resolutions and batch sizes

Changelog

v1.4 (2025-10-28)

• Final verification of repository structure and model integrity
• Confirmed all file sizes and paths are accurate
• Validated YAML frontmatter format and HuggingFace compliance
• Documentation verified complete and production-ready

v1.3 (2025-10-28)

• Verified repository structure and model file integrity
• Confirmed YAML frontmatter compliance with HuggingFace standards
• Validated all file paths and sizes
• Updated documentation timestamp

v1.2 (2025-10-14)

• Fixed YAML frontmatter: removed base_model fields (these are base models, not derived)
• Streamlined tags to essential categories only
• Improved metadata compliance with Hugging Face standards

v1.1 (2025-10-14)

• Updated YAML frontmatter format (metadata now precedes version header)
• Optimized tag ordering for better discoverability
• Verified all model files and sizes

v1.0 (2025-10-13)

• Initial repository documentation
• Added SDXL Base checkpoint (6.94 GB)
• Added SDXL Turbo checkpoint (13.88 GB)
• Organized directory structure for checkpoints, diffusion models, and LoRAs

License

License: CreativeML Open RAIL++-M License

Stable Diffusion XL models are released under the CreativeML Open RAIL++-M license, which permits commercial use with the following key terms:

• ✅ Commercial use permitted
• ✅ Modification and redistribution allowed
• ⚠️ Use restrictions apply (see full license)
• ⚠️ Must include license and attribution

Key Restrictions: Cannot be used for illegal activities, generating harmful content, or violating privacy rights. See full license for complete terms.

Citation

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

@misc{podell2023sdxl,
    title={SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis},
    author={Dustin Podell and Zion English and Kyle Lacey and Andreas Blattmann and Tim Dockhorn and Jonas Müller and Joe Penna and Robin Rombach},
    year={2023},
    eprint={2307.01952},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

@inproceedings{sauer2023adversarial,
    title={Adversarial Diffusion Distillation},
    author={Sauer, Axel and Lorenz, Dominik and Blattmann, Andreas and Rombach, Robin},
    booktitle={arXiv preprint arXiv:2311.17042},
    year={2023}
}

Official Resources

• SDXL Base Model
• SDXL Turbo Model
• SDXL Documentation
• Diffusers Library
• SDXL Paper
• SDXL Turbo Paper

Contact & Support

• Issues: Report issues with models or documentation on Hugging Face Discussions
• Community: Join Hugging Face Discord for community support
• Repository: This is a local storage repository - for upstream issues, see official model pages

---

Repository maintained locally | Last updated: 2025-10-28 | Version: v1.4