Outpainting with Stable Diffusion on an infinite canvas.

Start with init_image:

Start with text2img:

It is recommended to run the notebook on a local server for better interactive control.

The notebook might work on Windows (see this issue #12 for more information) and Apple Silicon devices (untested, check guide here: https://huggingface.co/docs/diffusers/optimization/mps).

This project mainly works as a proof of concept. In that case, the UI design is relatively weak, and the quality of results is not guaranteed. You may need to do prompt engineering or change the size of the selection box to get better outpainting results.

Pull requests are welcome for better UI control, ideas to achieve better results, or any other improvements.

setup with environment.yml

git clone --recurse-submodules https://github.com/lkwq007/stablediffusion-infinity
cd stablediffusion-infinity
conda env create -f environment.yml

if the environment.yml doesn't work for you, you may install dependencies manually:

conda create -n sd-inf python=3.10
conda activate sd-inf
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
conda install scipy
conda install -c conda-forge jupyterlab
conda install -c conda-forge ipywidgets=7.7.1
conda install -c conda-forge ipycanvas
conda install -c conda-forge diffusers transformers ftfy
pip install opencv-python

For windows, you may need to replace pip install opencv-python with conda install -c conda-forge opencv

Note that opencv library (e.g. libopencv-dev/opencv-devel, the package name may differ on different distributions) is required for PyPatchMatch. You may need to install opencv by yourself. If no opencv installed, the patch_match option (usually better quality) won't work.

conda activate sd-inf
huggingface-cli login # ignore this if you have already logged in
jupyter lab
# and then open stablediffusion_infinity.ipynb and run cells

This should get you started without needing to manually install anything, except for having an environment with Docker installed and an Nvidia GPU. This has been tested on Docker Desktop on Windows 10 using the WSL2 backend.

First, update the .env file with your Huggingface token from https://huggingface.co/settings/tokens

Open your shell that has docker and run these commands

cd stablediffusion-infinity
docker-compose build
docker-compose up

Watch the log for the url to open in your browser. Choose the one that starts with http://127.0.0.1:8888/

Once in jupyter lab, run the noteboook "stablediffusion_infinity.ipynb"

• Troubleshooting on Windows:
  • #12
• False positive rate of safety checker is quite high:
  • #8 (comment)
• What is the init_mode
  • init_mode indicates how to fill the empty/masked region, usually patch_match is better than others
• The GUI is lagging on colab
  • It is recommended to run the notebook on a local server since the interactions and canvas content updates are actually handled by the python backend on the serverside, and that's how ipycanvas works
  • colab doesn't support the latest version of ipycanvas, which may have better performance

The code of perlin2d.py is from https://stackoverflow.com/questions/42147776/producing-2d-perlin-noise-with-numpy/42154921#42154921 and is not included in the scope of LICENSE used in this repo.