FastSD CPU offers a significantly accelerated Stable Diffusion experience on consumer hardware, targeting users who want to run advanced AI image generation locally without requiring powerful GPUs. It leverages Latent Consistency Models (LCM) and Adversarial Diffusion Distillation (ADD) to achieve rapid inference times, with multiple interface options including a desktop GUI, WebUI, and CLI.

How It Works

The project achieves its speed by integrating Latent Consistency Models (LCM) and Adversarial Diffusion Distillation (ADD), which allow for fewer inference steps. A key enabler is its robust OpenVINO support, optimizing models for Intel CPUs, integrated GPUs, and NPUs. This approach bypasses the need for dedicated high-end GPUs, making advanced diffusion models accessible on a wider range of hardware.

Quick Start & Requirements

• Install: Clone the repository and run install.bat (Windows), ./install.sh (Linux), or ./install-mac.sh (Mac).
• Prerequisites: Python 3.10+ and uv package manager. OpenVINO is recommended for optimal performance.
• Resources: Minimum RAM varies by model: 2GB for LCM, 4GB for LCM-LoRA, and 11GB for OpenVINO models (reducible with Tiny Auto Encoder).
• Docs: Installation Guide, WebUI Screenshot

Highlighted Details

• Achieves 0.82-second image generation (512x512) on a Core i7-12700 using OpenVINO with SDXS-512-0.9.
• Supports multiple interfaces: Desktop GUI, WebUI (with Lora, ControlNet), and CLI.
• Offers real-time text-to-image generation (experimental).
• Integrates with Intel AI PC features, including NPU acceleration for Meteor Lake and Lunar Lake processors.
• Supports GGUF models via a stable-diffusion.cpp shared library.

Maintenance & Community

• Actively updated with new features and model support (e.g., SANA Sprint, Tiny AutoEncoder, Intel Lunar Lake NPU).
• Community support channels are not explicitly listed, but contributions are acknowledged.

Licensing & Compatibility

• MIT License.
• Permissive for commercial use and integration with closed-source applications.

Limitations & Caveats

• Real-time text-to-image is experimental.
• OpenVINO is not supported on Mac M1/M2/M3 chips (works on Intel Macs).
• Tiny Auto Encoder may not function in NPU mode.