AI Generated Fake Celebrity Nudes (not deepnude/deepfake)

Gargoyle521

New Member
Yes she's on the list. Reminder, I have over 500 names on this list, and it can take me a while to gather the images I need to create a model, so just because they're on the list doesn't necessarily mean I'll get to them. Every time I make a new model, I go through the whole list and do a sort of tournament, randomizing the order and picking a winner between each pair, and then doing that several times until I have a winner for who I most want to do next, but obviously that's going to be based on my own biases.

Daisy Ridley does seem like one that's more likely than some of the others to be picked at some point, but it's impossible to say when. I just do these one at a time. I've done 18 celebrities in about 9 months of training these models, and I think I've picked some good ones, but obviously I won't be able to do what everybody wants, so if any of you have a 12GB or larger GPU, I would encourage you to look into learning how to train your own models or at least LORAs.
Hey, you’re all good. I appreciate your work. You do you. I just wanted to see if she was on your radar at all. Thank!
 

fudefrak

Well-Known Member
Wasn't sure if I wanted to post this or not since she's not exactly a celebrity, but I did do a model for Cascina Caradonna. Youtuber/actress/musician, but most well known as being the face model for the character of Dina in The Last of Us 2. I'll put it in spoiler tags to avoid cluttering the thread.

00048-201145380.png00041-879168683.png00039-348240795.png00036-3910177451.png00035-3790890563.png00033-552768787.png00030-2797992918.png00028-3735644450.png00025-4099487832.png00023-747106478.png00021-857793404.png00018-598295327.png00015-295703200.png00009-1324279663.png00007-2497901326.png00004-2223457355.png00002-2277166128.png00013-2616786540.png00009-875343627.png00003-3797107546.png00001-258714252.png00029-1488730953.png00026-2251844265.png00016-2367654362.png00014-887769308.png00011-4002340251.png00007-4052715580.png00065-3523481781.png00061-2673834079.png00060-252634228.png00056-2847299249.png00053-3232979236.png00050-1733400947.png00049-3586078901.png00034-4051073498.png00016-3276354232.png00011-2829650820.png00007-4202067368.png00006-1348416739.png00003-909820392.png00020-2829272458.jpg
 

fudefrak

Well-Known Member
Hayley Williams

Honestly, this didn't turn out as well as I would have hoped. I probably needed more images to match the quality of some other the others, although I have done some models with similar number of pics that turned out better, so perhaps just with training more models and merging them together I can sort out some of the issues, but if not I do have some ideas to tweak my settings which shouldn't impact most celebrities with higher image counts, but may help a little for those with less. It's also just possible my dataset just may not have contained enough variety for the training to learn her face as well. It's nowhere near as bad as default Stable Diffusion, just not quite up to my standards. Although a few photos still looked pretty good.

00001-2565369247.png00004-481757896.png00007-751076286.png00009-994931769.png00012-4148479538.png00016-4071897752.png00020-909189121.png00022-2845694388.png00024-3339105079.png00030-1067467734.png00033-874588459.png00035-3567744557.png00037-2982434477.png00040-3472033397.png00042-3682051468.png00045-1092130129.png00050-2895098072.png00052-1802989757.png00055-894471758.png00058-690283196.png00047-3079559174.jpg00061-2402021449.png

Edit: Yeah, merging with a second model has seemed to improve the likeness quite a bit

00005-2735920996.png00008-4038729702.png00010-444350918.png00015-1690579604.png00023-1343678171.png00026-65398810.png00027-3167358392.png00030-504679051.png00034-1161646557.png00038-833224521.png00042-2302102312.png00046-1750942648.png00049-3574799599.png00035-1073801291.jpg00058-917324222.jpg00040-2043524964.jpg00055-2757836479.jpg00013-863092351.jpg00037-121308507.jpg00007-3609940021.jpg00004-3223311442.jpg00005-3061927046.jpg00014-223697375.jpg00018-777220708.jpg00031-905716643.jpg00054-3094671251.jpg00023-1100100262.jpg00016-1585346204.jpg00025-2917623649.jpg00021-2295337501.jpg00001-3348464352.jpg00017-2417943740.png00014-151295048.png00008-3986492435.png00005-672741001.png00003-638846968.png00034-536413931.png00032-1812702551.png00029-4104966543.png00023-2371991634.png00022-1720418645.png00020-2179125837.png00018-602132753.png00016-2232635887.png00013-601927287.png00007-2941801457.png00004-3875678068.png00001-2624843715.png00060-471080106.png00057-2911798148.png00055-84230024.png00054-3956610684.png00048-1968226767.png00046-1995145489.png00044-2209584981.png00037-3034594367.png00035-1503072650.png00033-57109015.png00052-766791465.jpg
 
Last edited:

fudefrak

Well-Known Member
I was recreating my Emma Stone model yesterday after , and I thought the results were a little underwhelming, so I decided to make the change to my training settings that I was considering earlier while training the Hayley Williams model, and the results turned out to be really good. You can see the new images posted after the edit in the

Basically, I was modifying the number of training epochs based on the dataset size, with the idea that the more images in my dataset, the more blending of concepts start happening, so you shouldn't need as many epochs, as long as the total number of steps still increases. So basically, my training steps were calculated by the square root of the number of images, multiplied by a constant value of 2700, plus 5 epochs for warmup. So with 1396 images included in the dataset for Emma Stone (673 images of Emma and 723 nude/porn images) that was coming to 77 epochs. I think the main issue is that I use a polynomial learning rate scheduler, starting at a high learning rate and getting smaller and smaller each step, almost like an exponentially decaying learning rate, focusing on learning smaller and smaller details, to really get those precise details, and the problem is you don't get through as many epochs before the learning rate drops considerably if the total epochs is 77. So while the number of training steps total does increase with a bigger dataset, the number of epochs was decreasing by too much when datasets got too big. So I just went ahead and reset it to the more typical default of 100 epochs, with the first 5 epochs being the warmup, and the results were much better. 30% more training time (which can be hours) but it was worth it.

Of course I'll be resetting my Emma Stone model again whenever high quality Poor Things screencaps come out anyway, lol, but this will still be useful for other new models I train, or further revisions on existing models.
 

lala2005

New Member
Emma Watson

I think my initial results for Emma Watson are fantastic, but I think I'm going to modify the captions in my dataset to make her short hair less likely to be generated. I personally prefer her having longer hair a lot more, and I know that's a somewhat common opinion, so if I go through and include "with short hair" in the images where she has short hair, and then retrain, it should result in a model that's less likely to automatically generate her short hair without me specifically including those words in the prompt.

Original model:

View attachment 2277109View attachment 2277110View attachment 2277111View attachment 2277112View attachment 2277113View attachment 2277114View attachment 2277115View attachment 2277305View attachment 2277306View attachment 2277307View attachment 2277308View attachment 2277309View attachment 2277310View attachment 2277311View attachment 2277312View attachment 2277313View attachment 2277314View attachment 2277315View attachment 2277316View attachment 2277317View attachment 2277318View attachment 2277319View attachment 2277320View attachment 2277321View attachment 2277322View attachment 2277323View attachment 2277324

tweaked the dataset. The new model does do the short hair less often but still does it on occasion without prompting, although none of the super short at least which I think looked the worst on her. There were some images in the dataset from red carpet appearances where it was difficult to tell whether she actually had short hair or it was just pulled back or in a bun or something, so I may still go back and see if I want to make any further changes, but any more models I make will be merged with this one:

View attachment 2281524View attachment 2281525View attachment 2281526View attachment 2281527View attachment 2281528View attachment 2281529View attachment 2281530View attachment 2281531View attachment 2281532View attachment 2281533View attachment 2281534View attachment 2281535View attachment 2281536View attachment 2281537View attachment 2281538View attachment 2282548View attachment 2282549View attachment 2282550View attachment 2282551View attachment 2282552View attachment 2282553View attachment 2282554View attachment 2282555View attachment 2282556View attachment 2282557View attachment 2282558View attachment 2282559View attachment 2282560View attachment 2282561View attachment 2282562View attachment 2305047View attachment 2305048View attachment 2305049View attachment 2305050View attachment 2305051View attachment 2305052View attachment 2305053View attachment 2305054View attachment 2305055View attachment 2305056View attachment 2305057View attachment 2305058View attachment 2305059View attachment 2305060View attachment 2305061View attachment 2305062View attachment 2305063
Hi buddy, I really loved this pics. I was hoping to know what u use to make this pics and how do u do it
 

fudefrak

Well-Known Member
Hi buddy, I really loved this pics. I was hoping to know what u use to make this pics and how do u do it
I train my own models. I gather hundreds of images myself, manually crop them, and manually give each one a caption. In the case of Emma Watson, I ended up gathering over 700 images of her. Then I also have over 700 nude/porn images all manually cropped and captioned as well, so that the model not only gets better at learning the person, but also gets better at learning how to generate images of nudity and sex.

I train with dreambooth and my own custom settings, generating a full model for each celebrity. In between new celebrities I often train new iterations of the model and average them together as well, which helps eliminate some of the unnecessary weights in the model and focus more on the changes common to each iteration which creates a better performing model.
 
Top