Over 1 year to create - this video combines my first 5162 attempts to speedrun Super Mario Bros. I recorded 193 hours of attempts (and practice) on an original 1985 nintendo entertainment system - then I wrote a custom computer program to process those videos and combine them via machine learning and conventional image processing techniques.

Support via ko-fi: https://ko-fi.com/flibidydibidy

*HERE IS A LINK TO THE 'How it's made / Director's commentary':    • The Unseen Side of Speedrunning - How...  

My final time was 4:58.941 That puts me in the top 150 on the leaderboard: https://www.speedrun.com/smb1 but hey maybe a sequel once I learn how to do 8-4 better and pull off the ol' bullet bill glitch :-).

I originally thought this wouldn't be that bad of a project: cool idea, couple months, done. But this challenged me in tremendous ways. It's done now, and it's cool to watch 193 hours of struggling (plus many, many times that developing, debugging, editing, and processing the videos) in just 5 minutes.

But this was not easy to do. For instance; when you have 41 million frames - even if something takes only one second - you're looking at well over a year of computer time alone. Most frames were processed in much less than a second, but some had to be done by hand to train the system and oh man I could go on forever on how it was made.

I've also have the idea to make this system work LIVE to combine multiple runners into a real race. Stick around if you're interested in that ;)