Skip to content

Latest commit

 

History

History
103 lines (68 loc) · 3.9 KB

README.md

File metadata and controls

103 lines (68 loc) · 3.9 KB
Raw

stablediffusiond

Pylint GitHub issues GitHub GitHub release (latest by date) GitHub Release Date

GitHub repo size GitHub code size in bytes Lines of code

A daemon which watches for messages on RabbitMQ and runs Stable Diffusion

  • Daemon - existing solutions use a webserver, here we use a daemon which is more lightweight
  • No hot loading - Model stored in RAM (10GB~) for faster processing
  • Less bloat - code and dependencies have been kept to a minimum
  • Flexibility - request daemon, response daemon and queue system can be run independently, allowing for more efficient use of resources
  • Easy to use - just run the daemon and send messages to the queue using send.py

img.png

  1. stablediffusiond listens to request queue
  2. client makes a request which is added to request queue
  3. stablediffusiond gets top item from request queue and runs Stable Diffusion based on request
  4. stablediffusiond places response from Stable Diffusion into response queue
  5. stablediffusion_responsed service listens to response queue and opens a socket on localhost:50007
  6. stablediffusion_responsed gets top item from response queue and returns it to client connected on localhost:50007

Requirements

  • RabbitMQ
  • Python 3.6+
  • conda (optional but recommended)
  • pika
  • Stable Diffusion and its dependencies
  • the Stable Diffusion Model

Installation

Run the installation script and follow the instructions

  1. curl -s https://raw.githubusercontent.com/w4ffl35/stablediffusiond/master/install.sh | bash
  2. Edit settings.py and fill it out with correct values for your environment.

Your directory structure should look like this:

> stable-diffusion
  > [various folders such as `assets`, `classes` and `scripts`]
  > stablediffusiond
    > [various files such as `receive.py` and `send.py`]

Manual installation

See installation script which contains comments for each step

Forked Stable Diffusion

Currently stablediffusiond depends on a forked branch which contains classes that mirror the code found in the txt2img and img2img Stable Diffusion scripts. The classes allow us to skip model initialization for each request.

The three classes can be viewed here (compare to scripts/txt2img and scripts/img2img):

Commands

stablediffusiond service

Starts a Stable Diffusion request queue runner

  • start sudo systemctl start stablediffusiond.service
  • restart sudo systemctl restart stablediffusiond.service
  • stop sudo systemctl stop stablediffusiond.service

stablediffusion_responsed service

Starts a Stable Diffusion response queue runner

  • start sudo systemctl start stablediffusion_responsed.service
  • restart sudo systemctl restart stablediffusion_responsed.service
  • stop sudo systemctl stop stablediffusion_responsed.service

bin commands

  • send a message to running stable diffusion queue stablediffusion_client '{"prompt": "cat", "seed": 42}'

Limitations

  • Installation is too manual
  • Very basic RabbitMQ configuration
  • Not production ready