stable-karlo

A Streamlit app that combines Karlo text-to-image generations with the Stable-Diffusion v2 upscaler in a simple webUI.

Now available on Google Colab:

Built with:

• Huggingface Diffusers🧨
• Pytorch
• Streamlit
• xformers

Install (Local)

Note that xformers is not in the requirements.txt. Using it is optional, but I'd recommend it if you have a GPU with low memory. You can follow the instructions on their repo to get it set up in the python environment.

Linux/MacOS

git clone https://github.com/kpthedev/stable-karlo.git
cd stable-karlo
python -m venv .env
source .env/bin/activate
pip install -r requirements.txt

Windows

git clone https://github.com/kpthedev/stable-karlo.git
cd stable-karlo
python -m venv .env
.env\Scripts\activate
pip install -r requirements.txt
pip install --upgrade --force-reinstall torch --extra-index-url https://download.pytorch.org/whl/cu117

Running

To run the app, make sure you are in the stable-karlo folder and have activated the environment, then run:

streamlit run app.py

This should open the webUI in your browser automatically.

The very first time you run the app, it will download the models from Huggingface. This may take a while, depending on your internet speed—the models are around 18GB total.

Memory Requirements

In the settings of each model, there are options for lowering the VRAM requirements:

• Both model settings have a Use CPU offloading option, which will substantially lower the VRAM usage.

• The Upscaler has two other methods to lower the VRAM usage:

  • Downscale input image - takes the Karlo image that is fed into the upscaler and downscales it to the specified size
  • Use xformers - uses xformers' efficent memory attention to lower the VRAM usage (this option requires a working xformers installation)

  ---

  Model	Optimizations	VRAM Usage
  Karlo	none	10GB
  Karlo	CPU-offloading	7GB

  Model	Optimizations	VRAM Usage
  Karlo + Upscale	none	>24GB
  Karlo + Upscale	Downscale to < 190px	12GB
  Karlo + Upscale	xformers	15GB
  Karlo + Upscale	CPU-offloading + xformers	15GB
  Karlo + Upscale	CPU-offloading + Downscale to < 190px	12GB
  Karlo + Upscale	CPU-offloading + xformers + Downscale to < 190px	10GB

License

All the original code that I have written is licensed under a GPL license. The licenses for the respective model weights, are included in the repository.

Changelog

• Dec 22, 2022 - Inital release
• Jan 02, 2023 - Add CPU offloading
• Jan 18, 2023 - Add Google Colab support
• Apr 06, 2023 - Update to use new Streamlit cache