On March 25th, OpenAI announced 4o image generation. And some guy had the genius idea to convert his family photos into the world of Totoro and Spirited Away.
Next thing I knew, my Twitter feed was full of this:
I hadn’t seen any NBA ones yet, so I figured I would give it a try. It was so easy.
Some of them failed in the classic missing appendages way, but if you point it out to the model it can usually fix it on the next try.
Other times it struggled with the surprising physicality of NBA players. Even in the cartoonish world of Miyazaki, Manute Bol and Muggsy Bogues are hard to believe.
But you can get there with a few nudges.
I dumped the full album of pictures that went well on imgur.
You might also enjoy The Auteur, The Cat, and The Flood or Grooves in latent space.
The most disappointing case was Drexler at the dunk contest.
He’s in such a unique pose, and the model just can’t get it’s head around it. Way out of the training distribution.
With some tweaking it gets worse…
… before it gets better. But never really good enough to capture the vibe of the incredible source material.
Lebron and Wade’s iconic shot in Miami is another one that just didn’t come out right no matter how much tweaking I did.
But at least it was kinda funny…