Spaces:
Running
dance the hully gully
This is a very interesting model and way more interesting than it looks
It isn’t limited to changing backgrounds and it's waay different than the ffmpeg type editors. It can do a ton. It’s strange in the way that the image and chat models are strange. You don’t know what you can’t do, so it’s possible you can do it. You dig new methods by exhaling them into existence. You can give it premises, like in improv comedy, and it takes you seriously and does it, at least part of the time. Here is a series of steps that might suggest a repeatable method for something the model can do.
I invented a structure, no idea if the model would do what I wanted. On the left is an empty, stage-like image, taken from SD 2.1 because I still think SD 2.1 is the best and most enjoyable. Then I have a prominent black line separating the sides. Then I have some raw material which doesn't have to be stored a certain way. The sky's the limit what the different things consist of.
prompt: please study the image and you should find a blue man in a black suit on the left side dancing the hully gully, the lady with the two long braids on the left side dancing the hokey pokey, a rusty metal anchor on the far right side, a checked futon on the right side. Now resize the weird ceiling to 1200x400 so it takes up most of the room.
It generates the world's first Dr. Manhattan/Mary Hartman Mary Hartman crossover. And they like each other, so we're all in big trouble.
Hand-crop to a square
Make a new two-part raw material image with the new square on the left and John Wilkes Booth on the right
prompt: please study the image and you should find a 19th century man with a moustache on the right side dancing the hoochie coochie, a rusty metal anchor on the far right side, a checked futon on the right side. Now resize the weird ceiling to 1200x400 so it takes up most of the room.
sic semper tyrannis!
Prepare a new two-part raw material image with the latest square on the left and Barbara Mason, soul genius, on the right
prompt: please study the image and you should find a blue man in a black suit dancing the hully gully, a lady with two long braids dancing the hully gully, a 19th century man dancing the hokey pokey, the cute black lady in the very long Van Heusen button shirt and no pants on the left side dancing the hully gully. Now resize the weird ceiling to 1200x400 so it takes up most of the room.
Video from various one-a-day websites (not from a HF space, too bad it is so much more locked down than it used to be)
Things I was varying as I went along:
say in the prompt that the new person is “on the right side” or “on the left side”. If one didn’t work, I tried the other, even if it didn’t match which side of the raw material the new information was on. Barbara isn't on the left side but I was tinkering until it worked.
whether to specify who is already dancing in the image or only mention the new person. I don't know.
Add a metal anchor way over on the right, literally as “an anchor”. Same with the futon. This was important for one of them. I think all the people and objects were bunched up over on the left until I forced it to spread out, or consider the whole frame as a canvas, by making it render an object way over on the right side. I also tinkered not using it later on and if there was enough stuff on the left and right, it wasn't needed,
Add more detail so that it would pick up who you were talking about and not combine features of two people into one, like Barbara Mason’s long Van Heusen shirt. (The model does not only know keywords about editing. It’s like an LLM.)
Why say “resize the weird ceiling”: the pixel counts are made up, and it isn’t really resizing anything. Resize the weird ceiling effectively made it erase the raw material area and make the full picture wider, which was terrific because then I could keep on cropping back to a square.
I hope you enjoy it.
Hello Nice i cant wait for see your characters dancing the " Mosh Potatoes "
