Easy-stabdif

This is a small wrapper script to run stable_diffusion.openvino with appropriate virtualenvs, possibly including the installation of some dependencies.

Usage

Requirements

Linux (Fully automated installation is currently only supported on Ubuntu Jammy, Ubuntu Focal and Debian Bullseye)
At least 16 GB of RAM.
A CPU supported by openvino

Quick start

./install.sh
./diffuse.sh --prompt "A beautiful meadow with many flowers, highly detailed oil painting" --seed 3735928559 --output meadow.png

Text to Image

The popular standard use case. You enter a description / prompt and you get an image back.

Use ./diffuse.sh. Documentation is available via ./diffuse.sh --help. Especially common parameters are:

--prompt [string]
Required. Used to specify what image should be generated.
--seed [integer]
Optional. Can be used to create reproducible output for the same prompt
--num-inference-steps [positive integer]
Optional. Lower values are faster to calculate, but tend to create fewer details and generally rougher / worse images. Default is 32. Experiment with values between 16 and 250.

Inpainting

Combines a (jpg) image and a (png) mask to modify parts of the original images while keeping other parts identical.

Use ./inpaint.sh image-basename unique-id prompt [additional diffuse.sh arguments].

./inpaint.sh connery my-first-inpainting "The most handsome man in the world: manly, beautiful, smart, well-dressed" --num-inference-steps 64 --strength 0.6 --seed 968734

This will:

Take image and mask with the basename "connery" (input/connery.jpg and input/connery-mask.png)
Combine them with the specified prompt
write the result to the file "output/connery-my-first-inpainting-a.png"
Perform a few additional inference steps without the mask to clean up possible edges
write the results of that operation to "output/connery-my-first-inpainting-b.png"

Especially common parameters are, in addition to those used in the Text to Image scenario:

--strength [float]
Optional. Lower values give less freedom to the neural net and remain closer to the input image

Image to Image

Takes a rough sketch and a prompt, and generates a new, detailed image from those.

Use ./image-to-image.sh image-basename unique-id prompt [additional diffuse.sh arguments].

./image-to-image.sh connery beardman "Sean connery with a large, bushy beard."