Table of contents

• nix-stable-diffusion
  • What's done
• How to use it?
  • InvokeAI
  • stable-diffusion-webui aka 111AUTOMATIC111 fork
• What's needed to be done
• Updates and versioning
• Acknowledgements

nix-stable-diffusion

Flake for running SD on NixOS

What's done

• Nix devShell capable of running InvokeAI's and stable-diffusion-webui flavors of SD without need to reach for pip or conda (including AMD ROCM support)
• ...???
• PROFIT

How to use it?

InvokeAI

1. Clone repo
2. Clone submodule with InvokeAI
3. Run nix develop .#invokeai.{default,nvidia,amd}, wait for shell to build
   1. .#invokeai.default builds shell which overrides bare minimum required for SD to run
   2. .#invokeai.amd builds shell which overrides torch packages with ROCM-enabled bin versions
   3. .#invokeai.nvidia builds shell with overlay explicitly setting cudaSupport = true for torch
4. Weights download
   1. Built-in way. Inside InvokeAI's directory, run python scripts/preload_models.py to preload main SD weighs and support models. (Downloading Stable Diffusion Weighs will require HugginFace token)
   2. Manual way. If you obtained SD weights from somewhere else, you can skip their download with preload_models.py. However, you'll have to manually create/edit InvokeAI/configs/models.yaml so that your models get loaded. Some example configs for SD 1.4, 1.5, 1.5-inpainting present in models.example.yaml .
5. Run CLI with python scripts/invoke.py or GUI with python scripts/invoke.py --web
6. For more detailed instructions consult https://invoke-ai.github.io/InvokeAI/installation/INSTALLING_MODELS/#community-contributed-models

stable-diffusion-webui aka 111AUTOMATIC111 fork

1. Clone repo
2. Clone submodule with stable-diffusion-webui
3. Run nix develop .#webui.{default,nvidia,amd}, wait for shell to build
   1. .#webui.default builds shell which overrides bare minimum required for SD to run
   2. .#webui.amd builds shell which overrides torch packages with ROCM-enabled bin versions
   3. .#webui.nvidia builds shell with overlay explicitly setting cudaSupport = true for torch
4. Obtain and place SD weights into stable-diffusion-webui/models/Stable-diffusion/model.ckpt
5. Inside stable-diffusion-webui/ directory, run python launch.py to start web server. It should preload required models from the start. Additional models, such as CLIP, will be loaded before the first actual usage of them.

What's needed to be done

• devShell with CUDA support (should be trivial, but requires volunteer with NVidia GPU)
• Missing packages definitions should be submitted to Nixpkgs
• Investigate ROCM device warning on startup
• Apply patches so that all downloaded models would go into one specific folder
• Should create a PR to pynixify with "skip-errors mode" so that no ugly patches would be necessary
• Shell hooks for initial setup?
• May be this devShell should be itself turned into a package?
• Add additional flavors of SD ?

Updates and versioning

Current versions:

• InvokeAI 2.1.3p1
• stable-diffusion-webui 27.10.2022

I have no intention to keep up with development pace of these apps, especially the Automatic's fork :) . However, I will ocasionally update at least InvokeAI's flake. Considering versioning, I will try to follow semver with respect to submodules as well, which means major version bump for submodule = major version bump for this flake.

Acknowledgements

Many many thanks to https://github.com/cript0nauta/pynixify which generated all the boilerplate for missing python packages.
Also thanks to https://github.com/colemickens/stable-diffusion-flake and https://github.com/skogsbrus/stable-diffusion-nix-flake for inspiration and some useful code snippets.