This repository provides a Gradio web UI demo for Stable Diffusion XL 1.0, targeting users who want to experiment with advanced image generation capabilities. It offers a user-friendly interface for generating high-quality images with features like a refiner model and multi-GPU support, simplifying the process of leveraging SDXL's power.

How It Works

The demo utilizes the Gradio SDK to create an interactive web interface. It loads both the base and refiner models for Stable Diffusion XL 1.0, allowing for a two-stage generation process that enhances image detail. The architecture supports data parallelism for multi-GPU acceleration and integrates optional features like Latent Consistency Models (LCM) LoRA and the faster SSD-1B model for improved performance.

Quick Start & Requirements

• Install dependencies: pip install accelerate transformers invisible-watermark numpy opencv-python safetensors gradio==3.11.0 git+https://github.com/huggingface/diffusers.git
• Launch: PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512 python app.py
• Requirements: PyTorch 2.0.1+, Python, GPU with CUDA.
• Resources: Requires significant VRAM for SDXL models.
• Links: Hugging Face Diffusers

Highlighted Details

• Supports Latent Consistency Models (LCM) LoRA by default for faster generation.
• Option to use SSD-1B (Segmind/SSD-1B) for even faster inference.
• Multi-GPU support via data parallelism (set MULTI_GPU=True).
• torch.compile support for potential inference speedups.
• Model offloading options (OFFLOAD_BASE, OFFLOAD_REFINER) to reduce memory usage.

Maintenance & Community

• Forked from the StableDiffusion v2.1 Demo WebUI.
• Actively updated with new features and model support.
• Encourages community stars and contributions.

Licensing & Compatibility

• License: MIT.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The demo's performance and memory requirements are substantial, necessitating powerful hardware. While torch.compile is supported, it adds initial compilation overhead. Model offloading saves memory but slows down generation.