Skip to content

Getting started

What you need

  • Node.js 22.19 or newer (nodejs.org; on Windows also winget install OpenJS.NodeJS.LTS).
  • For the built-in engine: an NVIDIA GPU is best (Apple silicon and the CPU work too, slowly), and about 4 GB of disk for the engine plus 5–7 GB per model.
  • Or a backend you already run: a server speaking the backend API, or AUTOMATIC1111 / Forge started with --api.

Support lists the systems, GPUs and browsers that are checked.

Install

git clone https://github.com/33rd-kk/latentry.git
cd latentry
./install.sh                                          # Linux, macOS
powershell -ExecutionPolicy Bypass -File install.ps1  # Windows

The script checks Node.js, installs and builds the web UI, and starts it. When it says Ready, open http://localhost:3000/setup in your browser.

Later, start Latentry again from the same folder with:

npm start

Ctrl+C stops it. Latentry listens on this computer only; to use it from a phone, see Other devices and security.

To check a release download before you run it, or to install without the built-in engine, see Install from source.

Set up the engine

The Setup page runs in three steps:

  1. Install the engine. Latentry looks at your GPU and installs Python 3.12, the matching PyTorch build and stock diffusers, all under .runtime/ in the app folder (about 3–4 GB, a few minutes). Deleting .runtime/ removes it again. The Python packages are the ones pinned for your machine, each file checked against its hash; Repair / update makes the engine's environment match them exactly, so a package you added to it by hand is removed. To take the newest packages instead (say, a pinned one has a known vulnerability and no Latentry update has come), set LATENTRY_ENGINE_PACKAGES=latest in .env.local and run Repair / update; Setup then says the packages are not pinned.
  2. Download a model. Each model comes from its publisher on Hugging Face. Read its licence (linked on the page; some are non-commercial), tick the box, and download. Or point the models folder at models you already have.
  3. Download a tagger (optional), for reading tags from pictures.

See Setup and the engine for what each part shows.

Prefer servers you already run? Add them on the Settings page instead (see Only the servers you already run).

Generate the first picture

  1. Open Generate. Pick the backend (engine for the built-in one) and the model.
  2. Write a positive prompt, for example 1girl, solo, smile, upper body, cherry blossoms.
  3. Press Generate. Progress shows per step; the pictures appear on the right and are saved to the gallery folder (output/ by default).

Next