Vintage and Retro Celeb Fakes (Stable Diffusion)

[Papyrus1000]

When you generate your images, do you use any actual nude images of these celebrities, where you can?

Not sure I get the question.

 

[sugarspicenodice]

Not sure I get the question.

he's asking are you using nude photos of celebs that you later modify or a different process all together?

 

[Youngla0450]

Not sure I get the question.

In other words, are you using those nude photographs of celebrities that exist, and generating new images based off those photographs? Or are you creating the images entirely by your own effort?

 

[Papyrus1000]

Sorry, I still find the question a bit confusing. It's like this: you teach a

You must be logged-in or registered for see links

to create the likeness of a celeb by training it on about 75-150 images.

You must be logged-in or registered for see medias

You must be logged-in or registered for see medias

If some of those source images are nude (including synthetic nude data, like, fake pics), you'll get a model that can create convincing NSFW content for that celeb.

REAL DATA (SOURCE FOR LORA):

View attachment View attachment View attachment

FAKE/SYNTHETIC DATA (ALSO SOURCE FOR LORA - MAY BE MORE THAN THE ACTRESS WAS WILLING TO DO IN HER HEYDAY):

View attachment View attachment


RESULT - LoRA THAN CAN DO MORE ACCURATE NSFW IMAGES OF THE ACTRESS (THESE ARE PURE TEXT-TO-IMAGE, BUT MODEL IS ALSO GOOD FOR INPAINTING REAL IMAGES):



Maybe the original question I didn't understand was wondering whether I do 'nudifying', like, taking a celeb image and removing the clothes, which I know is popular.

I don't do that. TBH the setup for most nudified pics is nuts, like Catherine Bach or someone else would be standing grinning stark naked at a red carpet event in 1982, and so on. About 94% of a convincing fake is a convincing context.

The other reason I have not used other people's nudified pics for my LoRAs is that the bodies are really cartoony and idealized, usually without pubic hair, and not looking like what the retro celeb would have looked like naked, back in the day. In the case of an actress such as Lana Wood, who actually has provided some real nude reference material, there's no excuse for training a 'generic' body, rather than the body she has already shown us.

 

[Papyrus1000]

It is very difficult to get Stable Diffusion to render female masturbation, so I have made a LoRA,

You must be logged-in or registered for see links

.

Source images were porn photos mostly from the 70s and 80s, no toys, dildos, etc, just hands.

I'm not putting it up on civit, though anyone else is welcome to upload it and take credit, or say they found it on the internet (don't credit me for it).




The trigger word is masturbatingwomanex, though it will probably work without it, or with similar words.
It is multi-trained, so the correct LoRA strength is around 0.15.

You are going to have to do a LOT of generations to get one where the hands aren't messed up. The better the celebrity LoRA, the less the hands will mess up, but the ratio is always going to be something like 1 in 20 at best.

Sometimes this LoRA will give you a nice 'presentation' without masturbation action, too.

 

[FilthyPervUK]

View attachment 2388446

Amazing work! Thanks for sharing.

 

[Shoestring]

It is very difficult to get Stable Diffusion to render female masturbation, so I have made a LoRA,

You must be logged-in or registered for see links

.

Source images were porn photos mostly from the 70s and 80s, no toys, dildos, etc, just hands.

I'm not putting it up on civit, though anyone else is welcome to upload it and take credit, or say they found it on the internet (don't credit me for it).

View attachment 2388445 View attachment 2388446 View attachment 2388448

The trigger word is masturbatingwomanex, though it will probably work without it, or with similar words.
It is multi-trained, so the correct LoRA strength is around 0.15.

You are going to have to do a LOT of generations to get one where the hands aren't messed up. The better the celebrity LoRA, the less the hands will mess up, but the ratio is always going to be something like 1 in 20 at best.

Sometimes this LoRA will give you a nice 'presentation' without masturbation action, too.

View attachment 2388447

Wow. Carrie was one of my first crushes. She was all I could think about when I came out of the pictures in 1977.
Now if you could do the same with Joanna Lumley as Purdey from the New Avengers, that would be amazing

 

[Papyrus1000]

Brigitte Lahaie (1976-80 era), trained on 261 porn images.

A new LoRA for the 1970s/80s French porn actress

You must be logged-in or registered for see links

. Trained three times in the usual way that I do, and needs a LoRA strength of 0.25 (this strength is pretty inflexible, because I merged at 0.7/0.7/1 and not the usual 1/1/1. so nothing else will work as well as 0.25).

You must be logged-in or registered for see links

You must be logged-in or registered for see links

File size is about 288MB, only tested with AUTOMATIC1111.

Have done very little with it yet, hardly any NSFW stuff, but will add images to this post when I can.

 

[Papyrus1000]

For anyone who'd like to see more of the

You must be logged-in or registered for see links

of Carrie Fisher as Slave Leia, I made a 'Slave Leia beach pinup' LoRA that can be used together with the Slave Leia LoRA/s (

You must be logged-in or registered for see links

/

You must be logged-in or registered for see links

).

I'm not uploading it to Civit, but you can

You must be logged-in or registered for see links

.




It uses all available existing real pictures from those shoots, with furries and Darth Vader etc., taken out, and also a number of cosplay photos on the beach. It is trained normally, so a LoRA strength of 0.5-1.0 should be fine. Trigger word is beachleia style, together with the trigger words for whichever Slave Leia LoRA you're using.

 

[Papyrus1000]

Here is a LoRA of British 1970s Playmate

You must be logged-in or registered for see links

- a simple Kohya conversion from a Dreambooth I did last year. LoRA strength is 1.0, and trigger is marilyncole woman.

You must be logged-in or registered for see links

.

You must be logged-in or registered for see links

.

LoRA file is 239mb.

 

[Papyrus1000]

A really basic LoRA of British early 1980s nude model and actress

You must be logged-in or registered for see links

, who later became the thriller author Mo Hayder.

Hardly done any NSFW yet.

This is a Dreambooth conversion to LoRA, and it's really as good as I was able to get it last year, considering how bad the early 1980s source material is. I'll post more NSFW of it as I get time.

LoRA is 297MB,

You must be logged-in or registered for see links

.

LoRA strength of 1 is fine, this is well, well below the standard of recent high-effort pieces like Slave Leia, but there if it interests anyone.

 

[Papyrus1000]

There is a new, better, multi-trained version of the

You must be logged-in or registered for see links

.

You can

You must be logged-in or registered for see links

(288mb).

You must be logged-in or registered for see links

Trained on a lot more synthetic data, as well as real photos, and is more flexible than the old version.

Haven't had a chance to do much with it yet.

 

[Papyrus1000]

There is a new, better, multi-trained version of the

You must be logged-in or registered for see links

(having failed to do a good LoRA V2 I tried again, and this went better).

You can

You must be logged-in or registered for see links

.

You must be logged-in or registered for see links

Needs a LoRA strength of 0.15.

Trained on a lot more synthetic data, as well as real photos, and is more flexible than the old version.

Haven't had a chance to do much with it yet. Will post more if I get time.

 

[Shoestring]

Wow. These are getting really good. If you ever get around to doing Lindsay Wagner they would be great

 

[Papyrus1000]

Wow. These are getting really good. If you ever get around to doing Lindsay Wagner they would be great

Sorry, always liked her, but not that way. Next one I'm doing now is Linda Evans (1979-82). After that probably Pamela Stephenson (of which I made a really crap DreamBooth last year, I think I can do better).

 

[Shoestring]

Never into Linda, but Pamela was fantasy of mine when I was but a young teen

 

[Papyrus1000]

Haven't even chosen which of the checkpoints to merge yet, but I think the Linda Evans LoRA could end up one of the most accurate and detailed of them all. Just quick tests at the moment, of the unmerged points...

 

[Papyrus1000]

So here is the Linda Evans LoRA, version 1,

You must be logged-in or registered for see links

.

You must be logged-in or registered for see links

.

You must be logged-in or registered for see links

.

Trigger is levansex woman.
LoRA strength needed is 0.15.

It has a few rough edges, but rewards effort.

 

[Papyrus1000]

Here is a new style LoRA for all-fours. I have included it in

You must be logged-in or registered for see links

at Civit that were previously only posted in this forum. I will also add later styles to this new collection at civit, for as long as my posts stay up there.

You can also download this all-fours LoRA at

You must be logged-in or registered for see links

.
Trigger is acarponex style.

You would think that the triple-trained more recent models would work better than the older ones, but it is not always so. Your mileage will definitely vary! Some of the oldest and simplest models work well with it, even if their source data did not contain any all-fours images.

Other old LoRAs, such as

You must be logged-in or registered for see links

, had a ton of all-fours images in the data, but don't work well with this LoRA.

This one was trained 'straight', just once, so a LoRA strength of 0.5-1.0 is the usable range.

Here are some *raw* and unretouched gen samples (except for automatic retouching of face by

You must be logged-in or registered for see links

). Pretty much any output using this LoRA (or most of the other style LoRAs) will need extra work. These are super-rough examples.

This LoRA was trained on only 100 vintage images. Would probably be better to train it at a higher batch size and higher learning rate on about 1000+ images.

You must be logged-in or registered for see links

You must be logged-in or registered for see links

(I'm really surprised this worked, it is not a great LoRA)


Dana Gillespie (unpublished - another surprise, this LoRA is terrible)

You must be logged-in or registered for see links

Fiona Lewis (unpublished - another under-performing LoRA that is strangely amenable to All Fours, though it is patchy)

You must be logged-in or registered for see links

View attachment

You must be logged-in or registered for see links

You must be logged-in or registered for see links

You must be logged-in or registered for see links

You must be logged-in or registered for see links

(unpublished - a generally poor LoRA, but amenable to All Fours)

You must be logged-in or registered for see links

You must be logged-in or registered for see links

You must be logged-in or registered for see links

 

[Papyrus1000]

There is a new POV blowjob LoRA, available at the

You must be logged-in or registered for see links

, or at

You must be logged-in or registered for see links

.

It is pretty effective, and only a handful of even the older LoRAs fail to work with it.

It is a straight train, so you can use it normally. Often, a LoRA strength of 1 is just fine, particularly if the character LoRA is resisting its influence.

The trigger is povex style, but adding more obvious prompts like fellatio, blowjob, oral and eye contact can't hurt.

Here are some examples from my models. They're all raw and unretouched, unlike the usual posts here, with plenty of mistakes, but show the potential of the LoRA.

Madeline Smith

Nicola Bryant


Audrey Landers


Brigitte Lahaie


Carrie Fisher


Linda Evans


Linda Lusardi


Lynda Carter


Margaret Nolan


Marilu Henner


Sabrina Salerno


Samantha Fox



Sheena Easton


Sophia Loren


Stephanie Beacham


Sybil Danning