About

This application stack provides an evironment to create & manage a Stable-Diffusion inference server. It is composed of a native front end app & dockerised backend server to ensure ease of setup and maintenance.

Requirements

• Docker see: https://github.com/NVIDIA/nvidia-docker#quickstart

Linux specific requirements:

• libzmq / libzmq-dev / libzmq3-dev (dependant on distribution)

Installation

• Install all requirements
• Initialise the docker container with start_docker.sh / start_docker.bat
• Run stable-ui

Features

• text to image / image to image / outpainting pipelines
• Stable Diffusion v1, v2 & sdxl support
• CPU support
• custom module extensions
• vae, pickle & safetensor model support
• low VRAM Support
• weighted prompts using compel (++ and -- to emphasize keywords)
• LORA support (TODO)
• infinite canvas inpainting & outpainting (TODO)
• textual inversion training (TODO)

Usage

Adding models

To add models to stable-ui select tools -> import model and load your .ckpt / .vae files. Once imported a copy of the model is saved to the data/models folder. The original file can be deleted.

Configuring models

On model import you will be prompted to provide configuration for a given ckpt, if you'd like to reconfigure a model you can redefine configuration via the tools -> configure models menu.

Loading models & generating images

The image generation options only become available once a model has been loaded into memory, you can do this by importing a model and then loading it with tools -> load model to memory. Once loading is completed the txt2img and img2img toolbar will unlock and you can begin prompting the model server.

Custom Modules

Functionality of the application can be extended by loading custom modules, see data/modules/ReadMe.md for additional information, default modules include:

• stable-ui: SD V1 & V2 support for image/text generation scripts, runs the sd generation ZMQ server within docker, critical component and should not be removed

Optimisations

In order to speed up inference time and use resources more effeciently the following optimisations are in place:

Always enabled / CPU inference

• jemalloc memory allocation library is used inside the docker image for more efficient memory allocation (https://github.com/jemalloc/jemalloc)
• numactl is installed to ensure that inference is pinned across all available CPU cores (https://linux.die.net/man/8/numactl)

Controls

• Middle mouse button to move the main window canvas

Building

There is a github pipeline available for both linux & windows that runs on the master branch. Release artifacts are saved for each tagged version and published.

If you'd like to build locally you can follow below steps after cloning the repo and initialising submodules:

• Initialise submodules: cd src/ThirdParty && git submodule update --init --recursive && git submodule update --remote

It's assumed that you are using the CMake build system.

Linux:

Install supporting libraries:

• libzmq / libzmq-dev / libzmq3-dev (dependant on distribution)
• zlib

Building:

• Run the following build scripts:
  • build.sh -r compile thirdparty libraries followed by stable-ui gui and package docker container, -r flag is optional to enable release mode
  • deploy.sh Deploy docker changes and synchronise with compose-up
  • run.sh Run gui
• Build is stored in build/stable-ui-bin (statically linked)

Windows:

• Compiler: VSCC Visual Studio 17 2022

The build_release.bat script assumes the you have vcpkg installed in the root C:\ drive
Libraries can be installed using vcpkg in the same manner as they are set up in the pipeline:

git clone https://github.com/microsoft/vcpkg.git C:\vcpkg
C:\vcpkg\bootstrap-vcpkg.bat
C:\vcpkg\vcpkg install zeromq --triplet x64-windows
C:\vcpkg\vcpkg install zlib

Building

• ./scripts/build.bat -r to build & package application, stored in build/stable-ui-bin, -r is optional to enable release mode
• ./scripts/deploy.bat to deploy / update existing docker container
• ./scripts/run.bat to start application