Shoestring
Well-Known Member
These show really promising potential. I can't wait to see the finished pics.
Vintage and Retro Celeb Fakes (Stable Diffusion)
- Thread starter Papyrus1000
- Start date
Papyrus1000
Well-Known Member
I'm unlikely to finish these particular pics. Sometime when the current spate of training dies down, I'll actually play with these models a bit, but that's not imminent. The pics are just proof-of-concept stuff.
Shoestring
Well-Known Member
Awww. I was looking forward to Nichola Bryant.
Papyrus1000
Well-Known Member
Well, the model is pretty much garbage. That's one of the training chores to come that gets in the way of actually using this stuff.
Papyrus1000
Well-Known Member
So this is the last of the 'style' LoRAs for a while, back to characters. This is a POV hairy lesbian LoRA, which is
Trigger is lespovex, and it should be kept at the lowest strength that will still keep the cunnilingus concept. For some models, it will have to be 1.0, where the faces of the training data may interfere a bit, and require inpainting after. For most models, it will work at 0.6 or 0.7.
Had no time to do many celeb ones yet:
Here are the generic samples at Civit:
Unshaved lesbian POV porn is pretty much non-existent. I had to add pubic hair to about 57 of the 69 images, with Photoshop and then inpainting, before training. The hair usually comes out okay, though.
You must be logged-in or registered for see links
, or can be
You must be logged-in or registered for see links
.Trigger is lespovex, and it should be kept at the lowest strength that will still keep the cunnilingus concept. For some models, it will have to be 1.0, where the faces of the training data may interfere a bit, and require inpainting after. For most models, it will work at 0.6 or 0.7.
Had no time to do many celeb ones yet:
Here are the generic samples at Civit:
Unshaved lesbian POV porn is pretty much non-existent. I had to add pubic hair to about 57 of the 69 images, with Photoshop and then inpainting, before training. The hair usually comes out okay, though.
Last edited:
Papyrus1000
Well-Known Member
Getting a bit nearer with the Pamela Stephenson LoRA, but it's tricky, because in that period (1978-83), she looked like an inflatable doll anyway (very android-y, almost too perfect), so realism is hard to gauge. I'll keep at it...
EDIT: Wow, this is incredibly hard. This model has defeated me so many times, and I think it may again!
I've produced dozens of versions and merges, and nothing works, except occasionally by blind chance. At the moment I am gradually cutting down the number of source images from the initial 137, getting rid of the lowest-quality ones.
Trouble with that method, if you take it too far, is you get a model that creates great renders but can hardly do anything (like the original
She may be fondly-remembered but she is not well-recorded from her heyday period. In case anybody else should ever try this, this is what you have to work with:
GETTY
Under-served. Nearly all shots from Superman III are B/W, the color ones are poor quality + you only get Lorelei character. All other Getty shots are grinning social stuff and goofy photo-shoots of limited value. Nearly all are in B/W, though some are amenable to colorization at palette.fm.
What's wanted are HQ publicity shots looking to camera, as are available for
Ebay, vintage erotica, general archives, Instagram, etc.
A tiny number of poor quality scans of poor quality magazine cover reproductions. I have taken text off using Photoshop generative AI, but that doesn't improve the photo quality.
TV Shows
Not the Nine O'Clock News would be great if there was anything available other than late 1970s-quality BBC PAL videotape. As it stands, this material is useless, even if you had a Blu-ray copy, because there is no better source available, and the occasional 16mm film inserts have not been remastered, as far as I know.
The Professionals - She made three appearances, one or two of which have some usable material. There is HQ material available from the original 16mm footage (not as good as 35mm, but it's okay for the few available close-ups).
The New Avengers - Same situation as The Professionals, I think she did two appearances, I have long since done all I could with that material.
But if you want a model that generates pictures where she actually looks at you, you will need to use the eye-direction neural filter in Photoshop (plus much retouching) to try and make this happen.
Space:1999 - She did one episode, with many close-ups, but they blitzed the soft-focus filter on her. Original source is 35mm, which is ideal, but the soft-focus dials down the value of that. Nonetheless, gives you some options.
Movies
Bloodbath at the House of Death - This is the best source available, with varied angles and multiple close-ups, and available on Torrents at super-high resolution, and not too much soft focus. The one nude scene is unhelpful and fleeting, but the diverse face angles are very useful. Still gives you the same eye-direction issue to solve manually.
Superman III - The soft-focus that Geoffrey Unsworth originally set up for this series is laid on thick here, and there is much less usable material than you might remember. Good for some profiles, but not as good as Bloodbath at the House of Death.
Les Patterson Saves The World - A bit after her prime period. I cannot find a HQ Torrent, and am not sure there is a Blu-ray.
History of the World: Part I - Useless period costume fodder, poor-quality cinematography.
It's possible that the only way to get a good model out of the available data is to extensively adulterate it all, using Controlnet to inpaint eyes and bodies directly, and extensive use of Photoshop neural filters. It's a 'month of Sundays' prospect, that I don't know I can commit to.
Anyway, have given up on multi-training, and am currently trying a full-size source data single train with reduced number of images. Will keep trying and keep cutting down the number of images, and may eventually get a workable model, but she will be pretty limited in what she can do, I think.
The only other thing I can think of is to use the semi-effective LoRAs I have just made to try and improve face/eye quality in the training data, and try again after.
EDIT: Wow, this is incredibly hard. This model has defeated me so many times, and I think it may again!
I've produced dozens of versions and merges, and nothing works, except occasionally by blind chance. At the moment I am gradually cutting down the number of source images from the initial 137, getting rid of the lowest-quality ones.
Trouble with that method, if you take it too far, is you get a model that creates great renders but can hardly do anything (like the original
You must be logged-in or registered for see links
model I did, almost entirely trained on Getty publicity shots, and which I have never been able to improve).She may be fondly-remembered but she is not well-recorded from her heyday period. In case anybody else should ever try this, this is what you have to work with:
GETTY
Under-served. Nearly all shots from Superman III are B/W, the color ones are poor quality + you only get Lorelei character. All other Getty shots are grinning social stuff and goofy photo-shoots of limited value. Nearly all are in B/W, though some are amenable to colorization at palette.fm.
What's wanted are HQ publicity shots looking to camera, as are available for
You must be logged-in or registered for see links
(
You must be logged-in or registered for see links
) and Marilu Henner (
You must be logged-in or registered for see links
)- but you only get these when an actress has been involved in a major US TV show (Taxi, Nanny and the Professor, etc.).Ebay, vintage erotica, general archives, Instagram, etc.
A tiny number of poor quality scans of poor quality magazine cover reproductions. I have taken text off using Photoshop generative AI, but that doesn't improve the photo quality.
TV Shows
Not the Nine O'Clock News would be great if there was anything available other than late 1970s-quality BBC PAL videotape. As it stands, this material is useless, even if you had a Blu-ray copy, because there is no better source available, and the occasional 16mm film inserts have not been remastered, as far as I know.
The Professionals - She made three appearances, one or two of which have some usable material. There is HQ material available from the original 16mm footage (not as good as 35mm, but it's okay for the few available close-ups).
The New Avengers - Same situation as The Professionals, I think she did two appearances, I have long since done all I could with that material.
But if you want a model that generates pictures where she actually looks at you, you will need to use the eye-direction neural filter in Photoshop (plus much retouching) to try and make this happen.
Space:1999 - She did one episode, with many close-ups, but they blitzed the soft-focus filter on her. Original source is 35mm, which is ideal, but the soft-focus dials down the value of that. Nonetheless, gives you some options.
Movies
Bloodbath at the House of Death - This is the best source available, with varied angles and multiple close-ups, and available on Torrents at super-high resolution, and not too much soft focus. The one nude scene is unhelpful and fleeting, but the diverse face angles are very useful. Still gives you the same eye-direction issue to solve manually.
Superman III - The soft-focus that Geoffrey Unsworth originally set up for this series is laid on thick here, and there is much less usable material than you might remember. Good for some profiles, but not as good as Bloodbath at the House of Death.
Les Patterson Saves The World - A bit after her prime period. I cannot find a HQ Torrent, and am not sure there is a Blu-ray.
History of the World: Part I - Useless period costume fodder, poor-quality cinematography.
It's possible that the only way to get a good model out of the available data is to extensively adulterate it all, using Controlnet to inpaint eyes and bodies directly, and extensive use of Photoshop neural filters. It's a 'month of Sundays' prospect, that I don't know I can commit to.
Anyway, have given up on multi-training, and am currently trying a full-size source data single train with reduced number of images. Will keep trying and keep cutting down the number of images, and may eventually get a workable model, but she will be pretty limited in what she can do, I think.
The only other thing I can think of is to use the semi-effective LoRAs I have just made to try and improve face/eye quality in the training data, and try again after.
Last edited:
Papyrus1000
Well-Known Member
Okay, I give up - 5-6 attempts are enough. I have put up a new
The single image below at the time of writing is the only one I have completed. The data had plenty of NSFW material, so presumably it can be forced to produce this too. If I succeed in doing anything like that, I'll add another post.
In spite of all the effort, this is the worst LoRA of the recent bunch - Stable Diffusion seems determined to turn her into Kim Basinger or Cheryl Ladd, who each have far more material in the underlying training data, and their identities seem to get entangled here.
So you may need to increase the strength of the model to get Stephenson to actually break through Stable Diffusion's huge repertoire of generic tarty blondes.
Good luck with it, if you're interested - you're going to have to really get out and push on this one! And I'm sorry I couldn't do better, I really tried.
Use a standard LoRA strength of 1.0, and it may well need more than that to get a likeness. Multi-training only made the problems with this character worse.
You must be logged-in or registered for see links
, and you can
You must be logged-in or registered for see links
.The single image below at the time of writing is the only one I have completed. The data had plenty of NSFW material, so presumably it can be forced to produce this too. If I succeed in doing anything like that, I'll add another post.
In spite of all the effort, this is the worst LoRA of the recent bunch - Stable Diffusion seems determined to turn her into Kim Basinger or Cheryl Ladd, who each have far more material in the underlying training data, and their identities seem to get entangled here.
So you may need to increase the strength of the model to get Stephenson to actually break through Stable Diffusion's huge repertoire of generic tarty blondes.
Good luck with it, if you're interested - you're going to have to really get out and push on this one! And I'm sorry I couldn't do better, I really tried.
Use a standard LoRA strength of 1.0, and it may well need more than that to get a likeness. Multi-training only made the problems with this character worse.
Last edited:
Papyrus1000
Well-Known Member
Since the Pamela Stephenson defeat, I've decided to give the celeb LoRA training a break and try and learn how to revive low-quality old magazine material. Those skills would have helped me a lot with Pam, and with many other retro LoRAs.
I am using a challenging and tiny set of porn images from 1979, the only ones available of this obscure model, it seems:
So the task is to upscale and inpaint, over and over again, using the SDV1.5 inpainting checkpoint, and any other tool that would help, and loads of ControlNet, and to just keep going until the resolution and quality are higher. I'm learning a lot, and will get better. Here is the first of the six.
The above image was too large to upload into a post. You can see the original res at
When all six are done, I will similarly upscale six generations from a DreamBooth model I did last year of this woman, and will use the 12 images to train a LoRA to make more training data. I just want to see if material this bad can really be made good enough for UHQ output...
EDIT - The second one is done:
Cannot upload full-size into this post - FULL SIZE VERSION IS HERE:
EDIT: Here are reduced versions of the final ones so far (it is too tedious to keep uploading the hi-res ones to image hosts)
And here is the last remake of the six original images - this was the hardest of all:
I am using a challenging and tiny set of porn images from 1979, the only ones available of this obscure model, it seems:
So the task is to upscale and inpaint, over and over again, using the SDV1.5 inpainting checkpoint, and any other tool that would help, and loads of ControlNet, and to just keep going until the resolution and quality are higher. I'm learning a lot, and will get better. Here is the first of the six.
The above image was too large to upload into a post. You can see the original res at
You must be logged-in or registered for see links
When all six are done, I will similarly upscale six generations from a DreamBooth model I did last year of this woman, and will use the 12 images to train a LoRA to make more training data. I just want to see if material this bad can really be made good enough for UHQ output...
EDIT - The second one is done:
Cannot upload full-size into this post - FULL SIZE VERSION IS HERE:
You must be logged-in or registered for see links
EDIT: Here are reduced versions of the final ones so far (it is too tedious to keep uploading the hi-res ones to image hosts)
And here is the last remake of the six original images - this was the hardest of all:
Attachments
Last edited:
Papyrus1000
Well-Known Member
Just taking a break from this long reconstruction project with...
A LoRA of 1980s 'Man with Two Brains;' actress
I just saw that she had been down in the list a long time, and checked out the available material. It was all pretty terrible, so I just pretty much threw it into Kohya SS and did a simple train, no multi-training, and it surprisingly is a pretty good LoRA, and more flexible than it should be.
I have other non-faking stuff that I need to get into now in this period, so I don't have time to work up good versions, and won't in the near future. But this LoRA is definitely capable of that kind of thing, and certainly of doing NSFW.
Here is some RAW output, with all the usual Stable Diffusion gotchas, just to give an idea:
RAW OUTPUT:
Don't know that I'll be here so much in the near future, but if I get time to return to the reconstruction project, will post progress.
A LoRA of 1980s 'Man with Two Brains;' actress
You must be logged-in or registered for see links
. Civit download is
You must be logged-in or registered for see links
, and alternate download is
You must be logged-in or registered for see links
.I just saw that she had been down in the list a long time, and checked out the available material. It was all pretty terrible, so I just pretty much threw it into Kohya SS and did a simple train, no multi-training, and it surprisingly is a pretty good LoRA, and more flexible than it should be.
I have other non-faking stuff that I need to get into now in this period, so I don't have time to work up good versions, and won't in the near future. But this LoRA is definitely capable of that kind of thing, and certainly of doing NSFW.
Here is some RAW output, with all the usual Stable Diffusion gotchas, just to give an idea:
RAW OUTPUT:
Don't know that I'll be here so much in the near future, but if I get time to return to the reconstruction project, will post progress.
Last edited:
sugarspicenodice
Well-Known Member
thank you for what you done so far and look forward to what you create in the possible future
Papyrus1000
Well-Known Member
There is a new, multi-trained version of the
Alternate download is
You must be logged-in or registered for see links
at Civit.
You must be logged-in or registered for see links
.Alternate download is
You must be logged-in or registered for see links
. | Thread starter | Similar threads | Forum | Replies | Date |
|---|---|---|---|---|
| M | Vintage erotica dot com | Off-Topic (Photos, humor, dating, etc.) | 17 | |
| F | Vintage starlets | Celebrity Old Leaked Photos and Nude Photos | 10 | |
| E | Emma photos vintage | Fake Celebrity Nude and Deepfakes | 0 | |
|
OnlyFans takinitvintage | OnlyFans | 1 | |
|
|
OnlyFans TheVintagePrincess | OnlyFans | 2 |