All Local AI Models in One Window — A Look at Uncensored Local Studio

If you've ever tried to set up a local AI stack, you probably know how messy it gets. You install Ollama or LM Studio for text, Automatic1111 or ComfyUI for images, a separate Whisper script for voice recognition, and yet another utility for text-to-speech. The result is gigabytes of Python virtual environments piling up on disk, conflicting ports, and VRAM getting eaten by multiple processes at once.

Recently I came across the techjarves/Uncensored-Local-Studio repository. The author tried to solve the problem of scattered utilities and bundled the main AI tools into a single desktop client.

What Uncensored Local Studio Can Do

This is a fully autonomous studio that runs locally on Windows, Linux, or macOS. The project has around 800 stars on GitHub and uses an open MIT license.

Inside the web interface, four tasks are combined:

1. Image generation and editing via Stable Diffusion (.safetensors, .gguf, and .ckpt files).
2. Dialogue with language models in GGUF format.
3. Real-time speech transcription from microphone or audio files.
4. Text-to-natural-speech conversion.

The main advantage is the complete lack of dependence on cloud APIs, external servers, registrations, or subscriptions. All computations happen exclusively on your computer.

How the Architecture Works

From a development perspective, the project interestingly combines existing high-performance C++ engines.

Instead of the usual heavy Python and PyTorch stack, the developer used compiled C++ backends. Image generation is handled by stable-diffusion.cpp, chat with text models by llama.cpp, voice transcription by whisper.cpp, and text-to-speech by kokoro-js with the Kokoro-82M neural network.

The frontend is written in Vite and React. The web server and backend lifecycle manager run on Node.js. So users don't have to spend time setting up an environment, the startup script downloads a portable version of Node.js directly into the app directory. System variables remain untouched.

VRAM management is solved interestingly. Chat and image generator work on a mutual exclusion principle. When switching tabs, the active engine releases video memory, allowing the app to run adequately on GPUs with 6–8 GB of VRAM.

Hardware Acceleration and Working with Models

The app can automatically detect installed hardware and selects the optimal backend:

• NVIDIA: CUDA optimizations are used.
• AMD and Intel Arc: computations run through Vulkan API or ROCm.
• Apple Silicon: the Metal engine is used (Intel Mac processors are not supported).
• Intel Core Ultra: there is experimental model running on integrated NPUs via OpenVINO.

For Windows, launching comes down to running windows.bat. On Linux and macOS, command scripts linux.sh and mac.sh are used.

Here's what the app directory structure looks like:

Uncensored-AI-Studio/
├── windows.bat                # Скрипт запуска для Windows
├── linux.sh                   # Скрипт запуска для Linux
├── mac.sh                     # Скрипт запуска для macOS
├── scripts/                   # Конфигураторы бэкендов и веб-сервера
└── app/
    ├── frontend/              # Исходный код интерфейса (Vite + React)
    ├── models/                # Модели Stable Diffusion (.safetensors, .ckpt)
    ├── llm-models/            # Текстовые модели (.gguf)
    ├── speech-models/         # Модели Whisper (.bin)
    └── outputs/               # Сохраненные генерации и метаданные JSON

The built-in model manager simplifies downloading: you can paste a direct Hugging Face link, and the file downloads straight to the right folder. Out of the box, it offers a quick start with lightweight models like Qwen2.5 Coder for chat or DreamShaper 8 for art generation.

The interface displays a built-in resource monitor showing CPU load, system RAM, GPU, and VRAM usage in real time.

Limitations and Caveats

Compactness has a downside. If you're used to specialized tools like ComfyUI with their node graphs, ControlNet, LoRA, or complex pipelines for Flux and Hunyuan, this project won't replace them. Only standalone SD 1.5 and SDXL checkpoints are supported here. Additional VAE files or text encoders are not loaded separately.

The second point concerns Linux systems. Pre-built backend binaries are compiled for relatively recent distributions with glibc 2.38 or newer (for example, Ubuntu 24.04). If running on older distributions like Ubuntu 22.04, the backend will throw a dynamic linker error. You'll need to either update the OS or compile binaries from source using the included script scripts/build/build_from_source.sh.

Who Will Find This Project Useful

Uncensored Local Studio is a great find for anyone who needs a universal personal assistant and image generator on one PC without the hassle of setting up Python environments.

For developers, the project is interesting as a clear example of integrating llama.cpp and stable-diffusion.cpp into a unified Node.js/React interface. The code is open, and the architecture demonstrates how to build responsive neural network applications without gigabyte-sized dependencies.