Videogen

Some notes from the videogen project I did for a few weeks in 2023:

Videos: https://www.youtube.com/@paramsingh258/shorts

thoughts:

• the videos were slop. the best image to video model i had available for actual use programatically was stable video diffusion and it only output ~4-5s for each image at very low resolutions.
• eventually i gave up on image to video completely as the videos were too low res for the algorithm to like
• shorts are better than full videos if you want algorithmic pull.
• subtitles are very important. shorts with subtitles do much better than shorts without them.
• the amount of views you need to make money off of youtube is kind of insane. as far as i understand, youtube doesn't count views that come from the shorts feed or from the home page feed for monetization. videos need to be really really good to make it.
• another thing i tried was a video with much more manual interference. i tried making a rk narayan story in the style of paperman. sadly, while it turned out reasonably, it took me a lot of time and i didn't really enjoy the process enough to put more time into it.
• overall, my personal intuition is thinking about this as an artist is better than thinking about it from a swe perspective. inevitably someone will use these models to make great art, it would be nice to be that person.