AI Generated Fake Celebrity Nudes (not deepnude/deepfake)

[Thall]

Charisma carpenter

 

[fudefrak]

Charisma carpenter

That's a deepnude. Read the title. There are other threads for that.

 

[fudefrak]

Yvonne Strahovski

 

[bleepedyblorpedy]

Can anyone out there do a "Mad Men" generated set? I think it could be amazing but I have no wherewithal or skill in making these! The January Jones set was so great it would be good to see the rest of the hot women (Elisabeth Moss, Abigail Spencer, Christina Hendricks, Alison brie...etc!) anyone feel like a challenge?!

 

[fudefrak]

The next celebrity I will be training a model from is Zendaya

 

[fudefrak]

Yvonne Strahovski

View attachment

 

[fudefrak]

retrained my Emma Stone model, including reference images from Poor Things, as well as a few other improvements to the dataset. Added some new pictures to the bottom of the Emma Stone 2 post, including a few images simulating if Poor Things went extra mile

 

[Diirge]

Couple of ai fakes of katy perry and megan fox

 

[The Grand Master]

Maisie Williams

The AI made her quite ugly lol

What the fuck is wrong with you. Generating images that makes Maisie look like a child. You need to fucking burn in hell

 

[fudefrak]

What the fuck is wrong with you. Generating images that makes Maisie look like a child. You need to fucking burn in hell

I didn't think any of them did (or I wouldn't have posted them), but apparently someone did so that post is gone now. Doesn't bother me since it was an old method that wasn't good anyway. I've considered deleting all the old method posts anyway, since the new method is actually giving me good results. I just left them up for historical purposes until I replace them with a trained model.

 

[fudefrak]

Dove Cameron

 

[fudefrak]

Emma Watson
View attachment

 

[Bobbyravioli]

Natalie Portman

 

[Bobbyravioli]

Emma Watson

 

[Bobbyravioli]

Alicia Vikander

 

[Bobbyravioli]

Kristen Bell

 

[Bobbyravioli]

Hilary Duff

 

[fallous19]

These are atrocious ffs, it literally butchers the faces on every set of photos iv seen..

 

[Namecaller]

Not to mention the Barbie fannies?

 

[fudefrak]

My guess is those images likely used one of those models that takes one of the newer stable diffusion versions (2.x, XL) and tries to train it to allowing nudity. The problem is, those models are heavily censored unlike the 1.x models, so if you want to be successful, you will need a MASSIVE dataset of nude and porn images, very good captions for the images, and aggressive and lengthy training sessions. Lora also won't cut it.

These also don't seem to be trained to focus on the celebrities being generated here, so they have a poor likeness, unlike my models which were trained specifically on each individual celebrity, as well as a large collection of nude photoshoots and porn.

In my experience, the newer SD models are going to be more trouble than they're worth, until I'm able to get a more powerful GPU and increase the size of my nudity datasets. While the newer models can render higher resolutions with less need for upscaling, it's just going to take a LOT more work to train them to doing what we want with them to a quality level that is satisfying, so I'd rather focus on the older models and using careful ControlNet parameters to upscale the results. I specifically include a lot of closeup shots in my datasets because I think that will help with the details when rendering larger resolutions. Without those closeups, the model may not have a clear idea of what certain features look like at high resolutions. I do need to add some more closeups to focus more on hands and open mouths though I think.

That being said, I keep all of my dataset images in their highest resolution possible, so whenever I end up upgrading to a new GPU (could still be years away) I'll be able to start training with the new generation of models, and come up with training parameters that are aggressive and lengthy enough to work.