A Tkinter app which uses Stable Diffusion inpainting to endlessly scroll through dynamically generated content. The image is continuously extended using inpainting.

• CUDA-enabled GPU capable of running Stable Diffusion.
• requirements.txt for Python requirements.

First, pull:

git clone https://github.com/zzbuzzard/stable-diffusion-infinite-scroll.git

and install the dependencies (probably in a virtual environment).

There are two scripts in this project: sd_scroll (for the scrolling app) and long_image (to generate very wide/high images using inpainting).

python sd_scroll.py [prompts] -s [number of steps] -d [H for horizontal or V for vertical]

Will start the full-screen scroller app, which will extend the starting image forever (until stopped by pressing Esc).

For example,

python sd_scroll.py "fantasy mushroom forest|peaceful landscape" -s 40 -d H

Note that prompts are separated by |, and surrounded by ". There are various other arguments (e.g. CFG, negative prompt, starting image), see --help for full details.

You can also choose the Stable Diffusion model to be used via the -m flag. By default, HuggingFace's runwayml/stable-diffusion-inpainting is used, but you could also use stabilityai/stable-diffusion-2-inpainting or any local model. I recommend using a specialised inpainting model.

python long_image.py [prompts] -s [number of steps] -d [H for horizontal or V for vertical] -out [output path] -n [number of shifts]

Most arguments are shared between the two scripts, see --help for full details.

For example,

python long_image.py "deep ocean life, deep sea photography, dark, jellyfish, bioluminescence, underwater, barren" -s 40 -d V -cfg 12 -out test.png -n 6

was used to generate the following 512 x 2048 images:

• A starting image can be provided. If not, it is generated using SD.
  • Note that the app will show a black screen until the first images are generated, see the console for details.
• If you run out of VRAM, try enabling the -as flag (attention slicing) or using a lower resolution (-r flag).
  • Everything was tested locally on a RTX 3060 Laptop GPU (6GB VRAM).
• Reducing the number of steps (-s) can lead to really fast scrolling.
• If you want to leave the scroller on for a long time, try using -spd with a value less than 1; this will cause the scroller to move slower but give the GPU regular breaks to prevent excessive heat.
• Vertical and horizontal scrolling are supported (-d H or -d V).
• Alternative models (including local models) can be specified with -m, but please be aware that this method works very poorly with non-inpainting models. If you get errors with local models, try passing --legacy.
  • XL models are supported: pass --xl at the command line as well as -m [YOUR_MODEL].

python sd_scroll.py "3D intricate structure, modern art sculpture, intricate, complementary colours, HD digital render" -s 40 -d H -cfg 12 -r 640 -as

I had to use -as for this one to reduce VRAM usage at the higher resolution of 640.

python long_image.py "beautiful rolling hills, peaceful landscape, blue sky, intricate detail, trees, sunrise" -s 40 -d H -cfg 12 -out test.png -n 10