AI Generated Fake Celebrity Nudes (not deepnude/deepfake)

[fudefrak]

Katy Perry

 

[Bobbyravioli]

Maisie Williams V2

 

[Shoestring]

Thanks for doing Maisie. Someone I've wanted to do for ages.

 

[Bobbyravioli]

Maisie Williams V2

 

[yuri_anzi32]

Katheryn WInnick

 

[Bobbyravioli]

Grok 2 is extremely good at nailing hard to generate celebs, but their nudity filter doesn’t allow fully nsfw images. Look how good some of these generations look, anybody know how to bypass or generate results this good without censoring?

In order:

Miranda Cosgrove
Charlize Theron
Zendaya
Hayley Atwell
Alexandria Ocasio Cortez (hot politician)
Blake Lively
Doja Cat
Anya Taylor Joy
Emily Ratajkowski

 

[LittleMouse1]

Grok 2 is extremely good at nailing hard to generate celebs, but their nudity filter doesn’t allow fully nsfw images. Look how good some of these generations look, anybody know how to bypass or generate results this good without censoring?

I feel like they all have the same body. Zendaya, Hayley Atwell, and AOC are all pretty different IRL.

 

[Bobbyravioli]

I feel like they all have the same body. Zendaya, Hayley Atwell, and AOC are all pretty different IRL.

It does a decent job at proportions, but I think the main standout is the faces. Haven’t seen a generator that gets them this close so consistently

 

[Bobbyravioli]

Few more that turned out pretty good -

Sydney Sweeney
Aubrey Plaza
Ella Purnell
Olivia Wilde
Jenna Ortega
Alexa Nikolas
Emma Stone
Jennette McCurdy
Olivia Rodrigo
Emmy Rossum

 

[LittleMouse1]

It does a decent job at proportions, but I think the main standout is the faces. Haven’t seen a generator that gets them this close so consistently

I guess I sorta maybe disagree with that.

 

[pseer_uk]

I guess I sorta maybe disagree with that.

They do look pretty good although I'm not sufficiently familiar with the subjects to make an informed opinion; which generator are you using?

I feel like they all have the same body. Zendaya, Hayley Atwell, and AOC are all pretty different IRL.

The issue I have with that statement is that the tits in all your examples are pretty much identical.

 

[Bobbyravioli]

I guess I sorta maybe disagree with that.

These are pretty great! Which generator did you use for these?

 

[LittleMouse1]

First off, apologies if I came of as being snarky up there.

I'm still using Stable Diffusion, run locally on my one and only work computer. Some of those models are 18 months old, which is "100 years" in AI right now, but it works great for people and stuff like that. Each person is trained separately. It's a lot of human time and then some amount of computer time to train a model. I have over 1000 models trained of people but I really only post all-generative stuff when I have time, and it takes me forever because I'm a purist and throw a lot away. I keep thinking I should do more and not worry, and I'm pretty sure the 4 or 5 pictures I did of Maisie Williams in a hurry to prove I can train small boobs are probably the ones I've gotten the most likes over.

 

[TimoFin]

Kate Garraway. AI. Hellraiser. Cenobite. Not for all tastes. Last one is of Kate and Lorraine, another series, NSFW because it's pure filth.

 

[Shoestring]

First off, apologies if I came of as being snarky up there.

I'm still using Stable Diffusion, run locally on my one and only work computer. Some of those models are 18 months old, which is "100 years" in AI right now, but it works great for people and stuff like that. Each person is trained separately. It's a lot of human time and then some amount of computer time to train a model. I have over 1000 models trained of people but I really only post all-generative stuff when I have time, and it takes me forever because I'm a purist and throw a lot away. I keep thinking I should do more and not worry, and I'm pretty sure the 4 or 5 pictures I did of Maisie Williams in a hurry to prove I can train small boobs are probably the ones I've gotten the most likes over.

Any good fake of Maisie will always go down well with me.

 

[pseer_uk]

Kate Garraway. AI. Hellraiser. Cenobite. Not for all tastes. Last one is of Kate and Lorraine, another series, NSFW because it's pure filth.

Great minds think alike Here's a little series I came up with; it's amazing what Grok lets you get away with as long as the participants are clothed

You must be registered for see links

You must be registered for see links

You must be registered for see links

You must be registered for see links

 

[fudefrak]

Sydney Sweeney

 

[fudefrak]

I noticed something while I was generating the above images, and it was too late to fix it for now, but I noticed that the way I named certain files in my dataset was causing faces to not be learned as well. Specifically, when I would have super closeup shots of a celebrity's face, where it was just their face, I would name it something like "A closeup photo of Sydney Sweeney's face on the red carpet for a movie premiere" or whatever. However, I noticed that I never use the word "face" when generating images. So it may be learning the concept of "Sydney Sweeney's face" as a separate idea compared to "Sydney Sweeney" on its own, and because of that, some of the more detailed aspects of her face may not be learned as well when not using the word face in the generation prompt. So I did a search through all my datasets and renamed files that were named that way to just be "A closeup photo of Sydney Sweeney ..." or whoever else had photos like this. Several of the celebrities were affected by this. Some more than others. Sydney had a lot of photos like that so I see some definite impact on her facial features on some of my renders here. As I regenerate these models, they should get better at learning their faces when I post new renders from them in the future.

For now, the next model I'm working on is Elle Fanning.

EDIT: Retrained the Sydney model and I do think it's doing a better job with her face on average. So that's encouraging for the other models this impacted.

 

[benderod]

I noticed something while I was generating the above images, and it was too late to fix it for now, but I noticed that the way I named certain files in my dataset was causing faces to not be learned as well. Specifically, when I would have super closeup shots of a celebrity's face, where it was just their face, I would name it something like "A closeup photo of Sydney Sweeney's face on the red carpet for a movie premiere" or whatever. However, I noticed that I never use the word "face" when generating images. So it may be learning the concept of "Sydney Sweeney's face" as a separate idea compared to "Sydney Sweeney" on its own, and because of that, some of the more detailed aspects of her face may not be learned as well when not using the word face in the generation prompt. So I did a search through all my datasets and renamed files that were named that way to just be "A closeup photo of Sydney Sweeney ..." or whoever else had photos like this. Several of the celebrities were affected by this. Some more than others. Sydney had a lot of photos like that so I see some definite impact on her facial features on some of my renders here. As I regenerate these models, they should get better at learning their faces when I post new renders from them in the future.

For now, the next model I'm working on is Elle Fanning.

Always appreciate your work and efforts
Really looking forward to seeing Elle Fanning, I've seen very few good or accurate edits of her, really underrated imo!

 

[fudefrak]

Always appreciate your work and efforts
Really looking forward to seeing Elle Fanning, I've seen very few good or accurate edits of her, really underrated imo!

It's looking like my dataset for Elle may even be as big as Sydney's, which surprises me, because Sydney was my biggest so far at 1028 images. With Elle I'm currently estimating 1049, but I've only gathered 323 so far, so we'll see.