Sºʟ. [Official AI-Generated Pixel Art Music Video]

Описание

This video was created using KaliYuga's Pixel Art Diffusion v3.0. P.A.D is the result of training the image model Disco Diffusion with a bunch of pixel art data to generate imagery using stylized pixel art techniques.

Stream Sºʟ. & ʜᴇᴀʟ̲+ʜ now!
https://linktr.ee/Acotu

Some thoughts:

This was my first effort at making an AI-generated video and I learned a lot. Many times I was shook by its complex & vivid output. It was only towards the end of the video that I felt like I had control over the composition, imagery, & movement of the frames

The best way I can describe this creative process would be repeatedly setting up a panda bear on a skateboard and pushing it in a certain direction. Sometimes the panda careens immediately and you need to recalculate, sometimes it coasts wonderfully for 10-15 seconds

Rendering each frame took 60s–90s, and the final video consists of ~2500frames. Post-processing in After Effects was used to display the realtime prompt & animation settings, plus adding noise grain/contrast & brightness/glow fx. The original output was 12fps, which was doubled in After Effects using frame blending. Low FPS + low resolution (640x320) = very quick results compared to similar methods of from-scratch AI video generation!

Google Colab is a great tool that lets you rent processing time from powerful CPUs and GPUs elsewhere. My computer is not equipped to handle this stuff so it was worth the 10ish bucks a month

PAD3.0 Google Colab notebook:
https://colab.research.google.com/github/KaliYuga-ai/Pixel-Art-Diffusion/blob/main/Pixel_Art_Diffusion_v3_0_(With_Disco_Symmetry).ipynb

My notebook for this video:
https://colab.research.google.com/drive/1eqZAITVhx_XBgFsViVKLaBwmbHOftjSG?usp=sharing

If you are interested in making your own Disco Diffusion images or videos, this reference was invaluable:
https://botbox.dev/disco-diffusion-cheatsheet/