>>62113137I first use Background remover in A1111 on my dataset which is filled with stream screenshots and the same-y background. This results in a bunch of pngs with transparent backgrounds which I place in a parent folder. Next to it I create a new folder called backgrounds where I add 300 backgrounds from danbooru. I go in the transparent image folder, add the script and run it.
It automatically merges the two images together with the backgrounds behind and the transparent image on top while keeping the resolution of the transparent image and making sure its entirety is filled with a background.
Job done! They can now be used in Lora training without baking the streamer's room.
>>62133326~Chuuu~
>>62133942>Someone had fun naming that or has been living under a rock kek.Apparently it's a math operation which represents an operation that can be executed only once and any subsequent applications do not change the result anymore. And their model allegedly does inference in a single step.
>This work was inspired by a Monologue from the sitcom Seinfeld. https://www.youtube.com/watch?v=glMHAXhPCFg Yep, this explains it well.
