Anime AI Companion Download Try the demo GitHub

Running an image model on your computer#

The pack studio can draw with a model that runs on your own computer: nothing is sent anywhere, nothing is paid per picture, and the cost is time and memory. The app does not bundle a model or Python. You start a small server, and the app talks to it on 127.0.0.1 only.

If you would rather not, pick another image maker in the studio's list: an API with your own key, or a website you already use through Chrome. This page is only for the first one.

What you need#

  • A computer with a lot of memory. The reference model needs about 22 GB of graphics or unified memory while it draws, and about 8 GB of disk. It was measured on an Apple M5 Max with 128 GB, and nothing else: a smaller machine may be too slow or fail.
  • uv, which fetches the Python packages when you run the commands below. Nothing is installed system-wide.
  • The app's own folder (the one you cloned), because the server is the file python/pack_studio/server.py in it.

1. Download the model once#

The reference model is FLUX.2 klein 4B (Apache-2.0). The server never downloads anything by itself, so fetch it first, into the Hugging Face cache:

uv run --with huggingface_hub python -c "from huggingface_hub import snapshot_download as d; d('black-forest-labs/FLUX.2-klein-4B', revision='e7b7dc27f91deacad38e78976d1f2b499d76a294', allow_patterns=['*.json','*.safetensors','*.txt','*.model','*.jinja'])"

2. Start the server#

From the app's folder:

uv run --with torch --with diffusers --with transformers --with accelerate --with sentencepiece --with pillow python python/pack_studio/server.py

It listens on http://127.0.0.1:7861 (--port changes it, --device mps|cuda|cpu picks the hardware). It loads the model on the first picture, so the first one is slow. Leave it running while you use the studio.

3. Tell the app#

Choose A model on this computer in the studio's Image maker list, or type:

/studio backend http://127.0.0.1:7861
/studio on

If the server is not running, or is running without the model, the studio says so and shows these steps instead of failing a picture.

Using another model#

Anything that speaks the same small protocol works in its place: GET /health, POST /edit and POST /generate, on 127.0.0.1. The protocol is described at the top of python/pack_studio/server.py. Give its address with /studio backend <address>. The model's own terms apply, and what it draws is your responsibility.

Improve this page on GitHub

Made with AI coding agents, mostly Claude Code. This documentation was written by Claude Code. Details