Skip to content

Latest commit

 

History

History
70 lines (45 loc) · 3.18 KB

README.md

File metadata and controls

70 lines (45 loc) · 3.18 KB
Raw

Infinite Zoom Stable diffusion

A colab notebook to generate infinite loop videos in minutes

it works on free colab plan

Open In Colab 👈 Easy to use notebook

built with gradio, with a nice UI

Open In Colab 👈 Advanced notebook

for more access and experimenting

automatic1111 webui extention 👈

Examples:

Dream of a Distant Galaxy:

Infinite.Zoom.-.Dream.of.a.distant.galaxy.online-video-cutter.com.mp4

Downtown Alley:
Infinite.Zoom.-.Downtown.alley.SHORT.mp4

Ancient Mechanism:
Infinite.Zoom.-.Ancient.mechanism.mp4

Colorful Fractals:
Infinite.Zoom.-.Colorful.fractals.SHORT.mp4

Cave of Skulls:
Infinite.Zoom.-.Cave.of.skulls.SHORT.mp4

Videos above were generated using the default settings in the notebook. Other than random seed only 1st part of the prompt was changed according to the video title (with slight alternations in some cases) e.g. prompt = "glowing colorful fractals, concept art, HQ, 4k".

Credits

  • Original idea and 1st version of the notebook was created by hardmaru
  • Thereafter BalintKomjati made the following improvements:
    • Adapted to run on Google Colab
    • Introduced "interpolation" between outpainted images to create smoother videos. This allows to use wider outpainting masks which tend to generate larger coherent structures, without unpleasant jumps between frames.
    • Put the whole product into a intuitive, user friendly notebook (at least that was the intention :) )
  • Then Vahid made the UI better and added a local version:
    • improved UI using Gradio
    • added some features including custom init image, multiple prompt feature, other models, etc.

Do you want to contribute?

Pull requests are welcome!

Here is a backlog of potential improvements to get you started :)

  • Add the possibility for the output to "drift" between prompts (now uses the same prompt for all images with different seed)
  • Fix the issue that any resolution higer than 512 will drastically decrease the network's imagination resulting the generated video rather boring. (e.g. upscale after each inpainting step?)
  • Make movement in the video more realistic without "fisheye" distorsions
  • Possibility to change movement of the camera during video (e.g. turns in any direction)
  • Generate video in non-cubic resolution (now only cubic is possible)