Skip to main content
Omnichar Studio is the open-source app. It runs on your own GPU with no account, and it’s free under GPL-3.0. Local generation and training cost nothing to run. Omnichar Studio canvas with model nodes wired together and a generated image in the output preview

​
What’s in it

  • A node canvas. Drop a model node, wire a prompt, and press Run. One node per model, with no loader, sampler or VAE wiring.
  • Characters. Build a .char from a photo or two and pick it from a dropdown on any node that takes one. See The .char character file.
  • LoRA training on your own images or short video clips, with the training nodes on the same canvas. The finished .safetensors lands in models/loras/, ready to use.
  • Local models: Z-Image Turbo, FLUX.1, FLUX.2, Krea 2, MiniMax H3 and LTX-2.5. See Supported models.
  • Hosted models through fal, with your own key. The key stays on your machine and each node shows the price before you run.
  • Extensions. Install community nodes from a GitHub repo. Nodes appear on the canvas without a restart.

​
Frames and takes

A frame holds every take you have generated for it. Generating again adds a take instead of overwriting the last one, so you can always go back to an earlier attempt. The take you pick as the hero is the one that flows into downstream nodes. Export zips a project into one archive with its inputs, outputs and the graph between them, so whoever opens it can re-run the pipeline. Video Director, Trim Video and Trim Audio nodes cut the results into a sequence.

​
How it runs

The Inline Core engine, written in Python, serves the web UI and runs every generation from one process on one port. You open the app in your browser. While it runs, the app serves its own API reference at http://127.0.0.1:8848/api.
The app has no login. If you start it with --listen, anyone who can reach the port can use it, the API included. Put it behind a firewall or a tunnel.

​
Next steps

Install

Install on Linux, macOS or Windows with one command.

Supported models

What each model does and how much VRAM it needs.