This project provides an uncensored, zero-setup, offline AI studio for Windows, Linux, and macOS, targeting users who need local, private AI capabilities without cloud dependencies. It unifies image generation, LLM chat, speech-to-text, and text-to-speech into a single, high-performance desktop interface, offering significant benefits in privacy and ease of use.

How It Works

The studio operates as a self-contained application, bundling Node.js, backend engines (stable-diffusion.cpp, llama.cpp, whisper.cpp, Kokoro TTS), and a web-based UI. It employs a mutually exclusive workspace design for text and image engines to optimize system resource usage. A key advantage is its automatic hardware acceleration detection, configuring and utilizing CUDA (Nvidia), ROCm (AMD), Vulkan (Intel/AMD/Nvidia), Metal (macOS), or OpenVINO (Intel NPU) backends without user intervention.

Quick Start & Requirements

Installation is initiated via platform-specific launcher scripts: windows.bat for Windows, linux.sh for Linux, and mac.sh for macOS. The Windows script automatically downloads a portable Node.js runtime and configures backends. Linux and macOS users must first make their respective scripts executable (chmod +x). A modern web browser is required. Linux prebuilt binaries necessitate glibc 2.38+ (e.g., Ubuntu 24.04, Fedora 40+); older systems may require OS upgrades or source compilation. macOS support is limited to Apple Silicon (M1 or newer); Intel macOS is unsupported. Optional setups for AMD ROCm and Intel NPU are available on Linux. A setup and demo video is available at: https://youtu.be/ESELhY-G_9w.

Highlighted Details

• 100% Offline & Private: Operates entirely locally with no internet, telemetry, or API key requirements.
• Zero-Install Portability: The entire runtime, including Node.js, models, and GPU backends, is self-contained.
• Auto-Configured Acceleration: Automatically detects and loads appropriate hardware acceleration backends.
• Integrated Model Manager: Allows downloading weights via Hugging Face URLs or importing local files.
• Live Performance Monitor: Displays real-time CPU, RAM, GPU, and VRAM utilization within the UI.
• Local Output Gallery: Saves generated images alongside their prompt parameters and metadata.

Licensing & Compatibility

The project is licensed under the MIT License, providing broad compatibility for commercial use and integration. It bundles stable-diffusion.cpp, also under the MIT License. Users must ensure compliance with the licenses of any downloaded model weights.

Limitations & Caveats

Intel-based macOS hardware is explicitly unsupported. The studio does not load companion files like LoRA, ControlNet, or VAE-only models as standalone image models. Linux users on systems with older glibc versions than 2.38 will need to upgrade their OS or compile backends from source. On dual-GPU Linux systems, manual selection of the desired Vulkan device may be necessary.