This project provides a fast, local AI image generation and editing solution designed for users with Apple Silicon Macs and NVIDIA GPUs, eliminating the need for cloud services or expensive hardware rentals. It offers state-of-the-art diffusion models with significant speed improvements and reduced memory footprints, making advanced AI image creation accessible on personal hardware.

How It Works

The system leverages multiple diffusion models, including FLUX.2 (klein variants) and Z-Image Turbo, optimized through 4-bit and INT8 quantization techniques (SDNQ, optimum-quanto) for lower VRAM usage and faster inference. It supports both Apple Silicon's Metal Performance Shaders (MPS) and NVIDIA's CUDA for cross-platform compatibility. FLUX.2 models enable text-to-image generation and image editing, while Z-Image Turbo focuses on rapid text-to-image generation, with a full version supporting LoRA adapters.

Quick Start & Requirements

A "1-Click" installation is available via Launch.command, which automatically installs dependencies in approximately 5 minutes. Manual installation involves cloning the repository, setting up a Python 3.11 virtual environment, and installing requirements with pip install -r requirements.txt. The web UI can be launched with python app.py and accessed at http://localhost:7860.

Highlighted Details

• Achieves fast inference times, e.g., Z-Image Turbo on an M2 Max generating a 512x512 image in ~14 seconds.
• Quantized models offer low memory requirements, with FLUX.2-klein-4B (4bit SDNQ) needing <8GB VRAM at 512px resolution.
• Supports advanced features like image editing with FLUX.2 models and LoRA adapter loading with Z-Image Turbo (Full).
• Provides cross-platform support for Apple Silicon (MPS) and NVIDIA GPUs (CUDA).

Maintenance & Community

No specific details regarding maintainers, community channels (like Discord/Slack), or project roadmaps were found in the provided README.

Licensing & Compatibility

The project inherits licensing terms from the original underlying models (FLUX.2, Z-Image Turbo). Specific usage terms and restrictions are not detailed within this README and require consulting the licenses of the respective base models. Compatibility for commercial use is not explicitly addressed.

Limitations & Caveats

The Z-Image Turbo (Quantized) model does not support image editing capabilities. The Z-Image Turbo (Full) model, while supporting LoRA, is noted as being slower. VRAM requirements vary significantly based on the chosen model and output resolution, with higher-quality or full models demanding substantial resources (e.g., 24GB+ for Z-Image Turbo Full).