A few years back, there was a site maintained by fan2000 known as the FAT WORLD wiki. Though it has long since gone offline, the idea of an alternate reality in which fatness (rather than thinness) was the central tenet of celebrity culture stuck with me, and I spent some time constructing the history of such a timeline for my own personal enjoyment. However, I’ve never been much of a visual artist, and the idea of simply running people through one of the numerous face filter apps out there didn’t seem all that satisfying to me. As such, the idea laid dormant for a while- right up until the recent advent of AI art. Suddenly it became much less difficult to get Photoshop-quality full body images of fat celebs, and I was reinvigorated by the ability to create visuals for my overall storyline.
The results have been released on my DeviantArt page this past couple of weeks, but now I’d like to extend an invite to anyone who might be similarly inspired by this alternate reality. I’ve set up an overall framework, fleshed out some initial pages, and written some guidelines for contributing- but a wiki is inherently a community website, so I can’t do it without you. So if any of this has sparked some inspiration in you, please check out BBWorld!
Stable Diffusion (my preferred implementation is SDNext), with a bit of touchup via paint.net.
I’ve worked with various SD models, but these pictures were all made using RealisticVision V4.0.
Thanks to everyone who’s contributed over the past few days!
Some recent highlights from the wiki:
Pompeia Paulina, by Hellosmith132:
Halle Berry, by Comicfan67:
And Paris Hilton, by yours truly:
could you post some prompts
To be fair, a large part of the process is just running through a lot of outputs to find something close to what you want, doing some inpainting to iterate on that (replace faces, improve hands, etc.), potentially doing some manual editing to emphasize particular body parts and feeding it back into img2img to smooth them out, etc.
But it is easier for some celebrities than others, which is likely some combination of being famous enough that the dataset can make a decent version of their face but not so famous that their body type is burned into their “core concept,” as it were. For instance, I was able to get this fairly quickly using the prompt “2009 closeup red carpet photo of morbidly obese and extremely fat Mila Kunis wearing a pink camisole” and a little bit of inpainting to fix a couple of facial features:
On the other hand, I was having trouble getting results with “fat pensive Yvonne Strahovski in The Handmaid’s Tale” until I switched models from RealisticVision to AbsoluteReality, and even then I had to include “wrinkles” in the negative prompt because the model doesn’t have a great idea of what she looks like; even so, although the image turned out okay, I think it’s pretty clear that Kunis is more recognizable:
So there’s really no golden rule for designing the ideal prompt- to some degree it’s luck, and to some degree it’s patience to work through a few iterations to progressively get closer to what you’re envisioning.