This page contains links to videos which were created using generative AI. Custom tools were written to generate these videos. Many different techniques were explored and the tools and patterns used evolved over time and continue to evolve.
The primary technique used for most of these videos is image to image diffusion pipelines. An image is generated, and then used to generate the following image. Modifications may be made between frames, such as as cropping and rescaling the prior image to create a zooming effect.
As only image diffusion models were used, these video have no temporal coherence. Unlike video models, where things happen in time, each frame is independent of all other frames except the one that comes before it. While this is limiting, certain styles still work well without temporal coherence, such as music visualization and psychedelic art. Video models usually would be preferred, but they are much slower and often too slow and limited on consumer hardware.
When this was created, I had no knowledge of other people doing this. I tried out many techniques, including using in-painting with various mask styles, and a whole lot of interpolation techniques. Only after I started uploading videos to Youtube did I find from Youtube's recommendations that there are other tools available for creating similar videos.
This page does not contain all the videos. To see all of the shared results, please see my Youtube Channel
This set of videos includes novelty songs. For details on what models were used, see the descriptions on Youtube. Songs were created with Suno, but lyrics were written externally. Most lyrics were created iteratively using GPT-4o. I played the role of the editor, and the LLM acted as the primary writer. For more songs, see the Youtube channel or this playlist.
These videos are zooms through different themes. There are many fractal zooms, some done with different mediums. A lot of these were created while experimenting with various techniques, and so the quality and styles vary. This is a small selection. See the Youtube channel for all videos.
Most of the earlier experiments didn't have any audio, but some of them still looked nice.
