techjarves/Portable-Local-Studio

What it is

Portable AI Studio packages four local AI workflows into one desktop web interface: Stable Diffusion image generation, GGUF chat via llama.cpp, Whisper speech-to-text via whisper.cpp, and Kokoro text-to-speech. Everything runs on the user's own machine with no cloud dependency, telemetry, login, or API keys. The project targets people who want a zero-setup, portable way to experiment with local models on Windows, Linux, or macOS rather than assembling Python environments and separate backends by hand.

The portability comes from bundling a Node.js runtime and precompiled GPU/CPU backends inside the project folder, so nothing is installed system-wide. An integrated Model Manager downloads weights from Hugging Face URLs or accepts drag-and-drop local files, and an output gallery saves generated images alongside prompt parameters and metadata. A live monitor shows CPU, RAM, GPU, and VRAM usage. Text and image engines are mutually exclusive by default to avoid exhausting memory, so users switch between workspaces in the UI.

Quick start

1. Download or clone the repository and keep it in a writable folder, since the first launch downloads runtime and backend files there.
2. Launch with the platform script:
   • Windows: double-click windows.bat. First run downloads a portable Node.js runtime and configures precompiled GPU/CPU backends. Windows 10 or 11 64-bit is required.
   • Linux: run chmod +x linux.sh, then ./linux.sh. NVIDIA users are prompted to set up CUDA; AMD users can add ROCm with ./linux.sh --max-perf; Intel Core Ultra NPU support is set up with ./linux.sh --setup-openvino.
   • macOS: run chmod +x mac.sh, then ./mac.sh. The prebuilt backend requires Apple Silicon (M1 or newer); Intel Macs are unsupported.
3. Add models by dropping .safetensors, .gguf, or .ckpt weights into app/models/, or download them through the Model Manager tab. Chat GGUF models go in app/llm-models/, Whisper models in app/speech-models/, and Kokoro assets in app/tts-models/ and app/tts-runtime/.
4. Open http://localhost:1420, select the model, and generate an image or start a chat. The setup and demo video is also available for a walkthrough.

Constraints

• Image models are single-file checkpoints only. SD 1.5, SDXL, and single-file SD/SDXL GGUF checkpoints load directly; Flux, HiDream, Hunyuan, Wan, Qwen Image, and Z-Image workflows are not supported because they need separate diffusion, VAE, and text encoder files. LoRA, ControlNet, VAE-only, text-encoder-only, and diffusion-only files are also not loaded as standalone image models.
• Linux binaries are built for Ubuntu 24.04-era systems and require glibc 2.38+ and GLIBCXX_3.4.32+. Older distributions may need an OS upgrade or a source build. OpenVINO NPU additionally requires kernel 6.6+, Python 3 with venv, and the Intel Linux NPU driver.
• The web UI has no login and only accepts connections from the local machine by default. Setting UAIS_HOST=0.0.0.0 exposes it to the network, and anyone on that network can then use it.
• Models carry their own licenses. The application code and the bundled stable-diffusion.cpp are MIT-licensed, but downloaded weights are subject to their creators' terms.

HotRepo's take

On 2026-11-01 the repository ranked 2nd on the JavaScript daily list and recorded 108 stars gained that day, reaching 1,691 stars and 415 forks. That is a single snapshot rather than a trend, but it places the project among the most-watched JavaScript repositories of that day. The pull is understandable from the feature set: it wraps four separate local inference stacks behind one launcher and one browser UI, removing the usual dependency wrangling that keeps many people from trying local models. The trade-off visible in the README is breadth over depth — single-file checkpoints only, no LoRA or ControlNet support in the image path, and platform-specific runtime floors. It suits users who want a working local playground quickly; users with advanced diffusion pipelines will still need other tools.

Project source and documentation