AI Generated Fake Celebrity Nudes (not deepnude/deepfake)
- Thread starter fudefrak
- Start date
Normally when I generate, I start with a 512x512 image (because that's how the models are trained) and then I feed that 512x512 image into img2img, with ControlNet as help to generate a 1024x1024 image. Some of these will have built in borders, as I've included borders on some of the training data images, so I can crop these images so that some of the results aren't exact square.
However, I wanted to experiment with something. I know trying to render at larger resolutions like 1024x1024 from the start tends to result in pretty unrealistic results, although the base model I use (RealisticVision 5.1) actually does an okay job with 768x768 sometimes, and therefore if I train enough models on the same dataset and merge them, the merge tends to do a decent job with that resolution as well, but the proportions tend not to be perfect regardless, so I haven't really been using that.
However, I thought of something. If I keep one axis at 512, and only increase the other axis, then the pixel mapping should mostly match the original images, especially when you think about closeups, which I do include a lot of in the dataset, both of the celebrity and of the numerous nudes I include.
So I did some tests tonight with a base resolution of 512x768, which mostly seemed to do a pretty good job, and then used my img2img ControlNet process to re-render these at 1024x1536, which also mostly turned out well. I don't think it's as consistently good as the normal process I use, but I can get full body results that match the level of quality I can get for topless-only shots in the past, which is to say much better looking faces and boobs. The bottom half tends not to be quite as good, but not terrible either, as I do have some closeups of pussies in my dataset, just not many closeups in poses like these.
Anyway, I decided to make a series of images like these for every celebrity I've trained a model for to give an example of what it can do:
I think the results are pretty decent for the most part. I think overall, the 1024x1024 based renders tend to be more consistent, and also take less time to render, but maybe on occasion I may do more like these as I do like the quality of them, although some of these models could use some improvement still. The tricky part is the png versions are often too big to upload here, so I will need to recompress them, which I never have to do with the 1024x1024 based renders. Perhaps I can improve my dataset to include better frontal closeups of pussies too, to help improve the quality of images like these in the future.
Btw, my next model will be ready soon. It's going to be Mila Kunis.
However, I wanted to experiment with something. I know trying to render at larger resolutions like 1024x1024 from the start tends to result in pretty unrealistic results, although the base model I use (RealisticVision 5.1) actually does an okay job with 768x768 sometimes, and therefore if I train enough models on the same dataset and merge them, the merge tends to do a decent job with that resolution as well, but the proportions tend not to be perfect regardless, so I haven't really been using that.
However, I thought of something. If I keep one axis at 512, and only increase the other axis, then the pixel mapping should mostly match the original images, especially when you think about closeups, which I do include a lot of in the dataset, both of the celebrity and of the numerous nudes I include.
So I did some tests tonight with a base resolution of 512x768, which mostly seemed to do a pretty good job, and then used my img2img ControlNet process to re-render these at 1024x1536, which also mostly turned out well. I don't think it's as consistently good as the normal process I use, but I can get full body results that match the level of quality I can get for topless-only shots in the past, which is to say much better looking faces and boobs. The bottom half tends not to be quite as good, but not terrible either, as I do have some closeups of pussies in my dataset, just not many closeups in poses like these.
Anyway, I decided to make a series of images like these for every celebrity I've trained a model for to give an example of what it can do:
I think the results are pretty decent for the most part. I think overall, the 1024x1024 based renders tend to be more consistent, and also take less time to render, but maybe on occasion I may do more like these as I do like the quality of them, although some of these models could use some improvement still. The tricky part is the png versions are often too big to upload here, so I will need to recompress them, which I never have to do with the 1024x1024 based renders. Perhaps I can improve my dataset to include better frontal closeups of pussies too, to help improve the quality of images like these in the future.
Btw, my next model will be ready soon. It's going to be Mila Kunis.
Mila Kunis
View attachment 
View attachment View attachment View attachment View attachment View attachment 
View attachment View attachment View attachment View attachment View attachment View attachment
Last edited:
Brie Larson
model still needs some work, particularly with her face, but will improve over time
Good timing with her new movie coming out today. That wasn't planned lol.
View attachment 
model still needs some work, particularly with her face, but will improve over time
Good timing with her new movie coming out today. That wasn't planned lol.
View attachment
Last edited:
Mr_Realistic
New Member
Hi new here let me post a few: Black Widow
Attachments
Last edited:
BigJohn3
Well-Known Member
I couldn't do anything close to this, so kudos to you but I was a bit put off with the fact that in the first few she looks quite like Peter Dinklidge
And as for what the hell is happening in picture 18???! That's some David Cronenberg stuff right there!
lol I considered omitting that one but forgot when uploading. AI mixes things up all the time when re-rendering in higher resolution. I tried multiple times with that one but couldn't get it to look better than this in higher resolution. I will eventually train more iterations of the Jenna model, and when I do additional iterations and average those models together, it tends to get better at weirdness like that. This was just the initial model so far.I couldn't do anything close to this, so kudos to you but I was a bit put off with the fact that in the first few she looks quite like Peter Dinklidge
And as for what the hell is happening in picture 18???! That's some David Cronenberg stuff right there!
Grlee
New Member
why do their faces look like that, such a demonic look
This was a much older AI tool that I haven't used in a long time. I kept those posts up for comparison purposes, but most of them are not great.
Grlee
New Member
what are you using now, and how can i get a hold of it, i have a couple of pics id like to try that AI on.
As explained in the first post, I now train my own models. I start with a variation of Stable Diffusion called Realistic Vision, and I train it using hundreds of images of the celebrity, plus hundreds of images of nude woman and porn. Every one of the images used in training has been personally gathered by myself. I either crop the images to a square, or shrink the image somewhat to fit inside a square, and I type my own descriptions of every image. Then training typically takes anywhere from 8 to 10 hours on my GPU, for a single iteration. I often go back and re-train these models, and average them with my existing trained models for even better accuracy. I try to gather at least 25 images every night for my dataset, so it can take a couple of weeks sometimes before I have a new model ready, but I'm always training new iterations of existing models every night if I don't have a new model to train yet.
Unfortunately, these models are 4GB each, so I won't be uploading them anywhere, but you're free to experiment with learning how to train models yourself. Beware, you really need a 12GB+ GPU in order to do what I do. There are some methods that use less, but the quality is not as good.
Also, note that using it on existing pics would be a "deepnude" which has a thread over here. I also use these models to make deepnudes from time to time.
Last edited:
