This lora creates the likeness of 1980's adult film star Traci Lords. It's not an exact likeness as it was cobbled together from a lot of OLD and vintage photos. Use it with your favorite realistic model at a strength of 0.8-1.0 For best results I would HIGHLY recommend combining this lora with the 80s-Babes lora https://civarchive.com/models/229489/80s-babes-vintage-pr0n-style Tips: I often use an anime model to get the pose i like and then remix into a realistic model like Photon. Combining traci with the 80s-Babes lora also works wonders. I always add "smiling, smile, happy" to my negative prompt and "pouting, pouty lips, sad expression" to my positive prompt to get that classic "Traci Pout". Good luck with it!
Shout out: This was a commissioned work created by Ggrue! He's the absolute best. Check out all his work here: https://civarchive.com/user/Ggrue
Description
FAQ
Comments (14)
use her full name as a trigger, and it blends with what sd1.5 knows her to look like and the results ain't bad at all....
..if you have an email I can shoot you a link to a 100% legal movie that I found that was upscaled. it was the best I could find that was legal, while scaping to make a lora. I couldn't find enough to go with it and gave up... but you might be able to use it, if you want to try again
Although this idea sounds good, it may no be necessarily the Best way
to train your model LORA or T.I. I have been doing alot of tests and creating
models since 2022 and I will explain to you the reasons why.
The original SD-1.5 program was trained with 1996 famous celebrity faces.
Many of the pictures used are of Stars when they look much older, some are
over 50+ years old.
By using the same name as the celebrity, you will get a mix of AI-created
photos from this data. If you create a younger looking "star" MODEL (LORA)
and use their "trigger word" as their real name, the AI will mix old faces
with your model and it will be difficult for the AI to create exactly what
you want to see.
Many months back some of us model creators discovered that the best way to
name "trigger words" for models is mix in some numbers into the name, so the
AI will NOT UNDERSTAND who the person is. It is best to NEVER name the model
of exactly what you are training your model on.
Here is an example - If your model is named "TracyLords", then you will replace
some letter with numbers and make the name totally UNKNOWN to the AI.
So , the new "trigger word" for the model will be "tr4c1_l0rds".
Or you can invent a totally fictional name which the AI will NEVER recognize
when training and creating the images for/from your LORA.
@maxhitman oh.. i'm not advising it... I'm just testing out his lora and seeing that using her full name works.
most of the time SD1.5 interferes and makes a mess... but this time it seems to be fusing wit the lora to give pleasant results...
... and to the lora creator, skip the email, just duckduckgo search for a 1080p upscaled version of 'tracey i love you' on a site that starts with Spank and ends with a Bang... wikipedia clearly states that she is over 18 in that movie
@JoeJo832 I didn't have great results using her name but maybe its just preference.
@jacklaughed positives... "traci lords, 18-years-old, (realistic photo:1.3), best quality, masterpiece, 8k, uhd,"
negatives.... "cartoon, two males, two females, asian, black and white, airbrush, smooth, blur, out of focus, grainy picture, bad face, distorted face, blurry face, deformities, bad anatomy, "
those are the standard copy-and-paste positives and negatives that I use for everything... except Asians and cartoons...
it looks like the lora poster used an image enhancer, she has a smooth look to her... to be expected with decades old photos found on the net... or screen caps...
...yeah, upon testing this a second time, I think it's worthy of 4 stars. She's not very "flexible" and distorts-to-hell rather easily if you give any prompting. there's no way she's cooking pancakes. Is that from too many pictures? I read 20 is the sweet spot... i'm still learning...
I think its just the image quality of old scanned images, they're not great source material for SD. It's def better than not having this lora at all.
Hey guys, thanks for all the interest. Here's a couple links for some of the images I've been able to pull off with this lora: https://i.imgur.com/tEaJegx.jpg https://i.imgur.com/DMWscTz.jpg https://i.imgur.com/pYuiQh5.jpg I often use an anime model to get the pose i like and then remix into a realistic model like Photon. Combining traci with the 80s-Babes lora also works wonders. I always add "smiling, smile, happy" to my negative prompt and "pouting, pouty lips, sad expression" to my positive prompt to get that classic "Traci Pout". Good luck with it!
Very cool and a good model. Thank you.
I will tell you a small trick that you can do , on certain faces which you may not have many photos to work with.
1 - First create a LORA and then create a large selection of images using checkpoints for realistic pictures, such as "Realistic vision".
Also use the original Stable-Diffusion 1.5 checkpoint for creating LORA and models. That way, you can be sure your model will also work best with all the other checkpoints created. Do NOT use cartoon or anime style checkpoints.
2 -Then pick out the AI-images which came out looking as much as possible to the real person.
Around 25 pictures is good enough.
Crop them into the usual 512x512 (or large) size
for creating another LORA. Select images of just the face, half-body portraits and a few for full size bodies.
3 - Once the second LORA is created, you will notice that your AI will create much better and realistic photos of the person. By having a realistic photo LORA created, it will also work well for creating cartoon or Anime-style faces too.
-
How to test your model LORA to see if they work at their best?
I will leave you with some other tips and tricks below...
-
If you have a great made model, but write bad prompts, you will not get the
best results from your created Lora or Textual-Inversion model.
-
Here are some useful tips and tricks for you to test and do.
-
- I will delete this post in 24 hours - copy everything I have posted for you.
This is only meant for us model creators to test their models.
-
For best AI renders....
Best Samplers:
Euler a (fast render but not always the best for face portraits)
DPM++ 2M Karras
DPM++ SDE Karras
------------------------------
CFG scale: best settings
Portraits: 7 - 8 - 9
Landscape: 9 - 12
-----------------------------
Steps: best settings
Portraits: 30 - 40 - 70
Landscape: 40 - 80 - 100
-
my personal choice for best quality AI photo-realistic photography
images is these settings ...
CFG scale - 8
Steps - 70
------------------------------
Depending on what Checkpoints you will decide to use, each checkpoint
will Create the images differently and with different effects on the images.
My personal choice for Photo-Realistic Photographs has always been...
- Your Main Checkpoint -> Realistic Vision version 1.3, or version 5.1
- Checkpoint Mergers...selected in your SD menu.
A - Protogen X34
B - Clarity version 1.9
C - Deliberate version 2
Overall,
this is the best mix of 4 checkpoints to use for Photo-Realistic AI pictures.
-------------------------------------------------
(BELOW) is a very good prompt script for doing tests and create
some nice images while you test your A.I. to create some famous faces.
Depending on what Checkpoint you will use, some times it will
create the face and body Very Good (A+), and
if it does not, then you may need to use a Lora, or a T.I. for training your
A.I. to make them better.
This is for a female NSFW photo-realistic Photograph !
I will also post for you a SFW image prompt below.
... Do not change the word placement, as this prompt has
been tested for many weeks as to be the best place for the words to be
in the script.
Each section has importance for the AI to understand better and to create better images.
...........................Positive prompt for Not Safe For Work .......................................
...IF ITS A LORA MODEL - You will write it like this ...
(<YOUR LORA MODEL:0.9>),a Realistic Full Portrait Photo of a ___place LORA Trigger word name here____ with long hair style, Beautiful, Detailed face, Perfect eyes, (highly detailed skin:1.1), (perfect breasts:1.1), ((perfect nipples:1.1)), perfect body, NSFW, nude, pussy, best quality, Modelshoot style, Professional Photography, Photorealistic, standing in a Dark Studio background, blurred background, RAW, analog style, sharp focus, 8k, ultra high resolution, DSLR, high quality, Fujifilm XT3, film grain, award winning, masterpiece
.
...IF ITS A TEXTUAL-INVERSION (EMBEDDING) MODEL - You will write it like this ...
a Realistic Full Portrait Photo of a ___place Embedding model name here____ with long hair style, Beautiful, Detailed face, Perfect eyes, (highly detailed skin:1.1), (perfect breasts:1.1), ((perfect nipples:1.1)), perfect body, NSFW, nude, pussy, best quality, Modelshoot style, Professional Photography, Photorealistic, standing in a Dark Studio background, blurred background, RAW, analog style, sharp focus, 8k, ultra high resolution, DSLR, high quality, Fujifilm XT3, film grain, award winning, masterpiece
.
...........................Negative Prompt for NSFW for all.......................................
(Clothes), (clothing), (pants), (bra), (underwear), (dress), (shirt), (skirt), canvas frame, (deformed iris, deformed pupils, semi-realistic, CGI, 3d, render, sketch, cartoon, drawing, anime, mutated hands and fingers, fused lips and teeth:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, doll, child, un-detailed skin, monochrome, black and white,
.......................................................................................
.......................................................................................
.......................................................................................
...........................Positive prompt for Safe for work .......................................
...IF ITS A LORA MODEL - You will write it like this ...
(<YOUR LORA MODEL:0.9>), a Realistic Full Portrait Photo of a ___place LORA Trigger word name here____ with long hair style, Beautiful, Detailed face, Perfect eyes, (highly detailed skin:1.1), perfect body, best quality, Modelshoot style, Wearing a Gold dress designed by Chanel, Professional Photography, Photorealistic, standing in a Dark Studio background, blurred background, RAW, analog style, sharp focus, 8k, ultra high resolution, DSLR, high quality, Fujifilm XT3, film grain, award winning, masterpiece
.
...IF ITS A TEXTUAL-INVERSION (EMBEDDING) MODEL - You will write it like this ...
a Realistic Full Portrait Photo of a ___place your embedding model name here____ with long hair style, Beautiful, Detailed face, Perfect eyes, (highly detailed skin:1.1), perfect body, best quality, Modelshoot style, Wearing a Gold dress designed by Chanel, Professional Photography, Photorealistic, standing in a Dark Studio background, blurred background, RAW, analog style, sharp focus, 8k, ultra high resolution, DSLR, high quality, Fujifilm XT3, film grain, award winning, masterpiece
.
...........................Negative Prompt for SFW for all.......................................
((NSFW)), ((Nude)), (Naked), (pussy), (vagina), canvas frame, (deformed iris, deformed pupils, semi-realistic, CGI, 3d, render, sketch, cartoon, drawing, anime, mutated hands and fingers, fused lips and teeth:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, doll, child, un-detailed skin, monochrome, black and white,
......................................................................................
IF YOU WISH for the model to be Outdoors, change the word script in the
section areas with these words below ....
Photorealistic, sunny, standing in a beautiful beach background, blurred background, RAW,
-or-
Photorealistic, sunny, standing in a foggy forest background, blurred background, RAW,
......................................................................................
.
.
Using the above posted prompt, your images will look much better in details
and in a Photo-Realistic quality (without the use of any extra Lora or T.I.).
--- This is good for training and testing your A.I. and Lora models to see if they are good.
.
The following images below were created just using the original Stable-Diffusion 1.5
without the use of any Lora/Lycoris/Textual-Inversion, and using the same style
of written prompt as I have posted for you above. In many cases your Stable Diffusion 1.5
has already been trained on creating over 1996 famous faces without the use of any pre-trained models.
It all depends on how well you write your prompts.
This is the full list of known faces used to train the original SD-1.5 program- https://laion-aesthetic.datasette.io/laion-aesthetic-6pls/celebrities
-
-
https://i.imgur.com/VuiQgOU.jpg
-
https://i.imgur.com/4CHOqDC.jpg
-
https://i.imgur.com/SuuGCpx.jpg
-
https://i.imgur.com/3jziP2r.jpg
-
https://i.imgur.com/XgdQVXX.jpg
-
- I will delete this post in 24 hours - copy the links I have posted for you
.......................................................
Definitely gets her amazing body right with those torpedo tits but having trouble getting a good likeness, Adetailer and inpainting aren't working great on her face (probably a result of low quality source material). Which checkpoint and sampler are you having the best luck with?
Hey there! I've been using Photon primarily but I ALWAYS pair this model with the 80s-Babe lora...it really adds that vintage style and feel. I use the model at full weight along with a special sauce of negative embeddings to boost quality. Good luck with it.
Oh I also try and use "happy, smile, smiling" in my negative prompt along with "pouting, pouty lips" in my positive prompts to get that classic Traci expression.
@MadripoolJoe Thanks for the tips on the expressions, I tried that lora but it appeared to be degrading the likeness a bit and the Traci lora gets her hair right anyway, so I'm fine without it. Are there specific negative embeddings you're using? The negative prompt would be helpful. Otherwise nice work, getting what you have with low-quality source material is impressive.
Hey everyone, thanks for all the interest. Here's a couple links for some of the images I've been able to pull off with this lora: https://i.imgur.com/tEaJegx.jpg https://i.imgur.com/DMWscTz.jpg https://i.imgur.com/pYuiQh5.jpg I often use an anime model to get the pose i like and then remix into a realistic model like Photon. Combining traci with the 80s-Babes lora also works wonders. I always add "smiling, smile, happy" to my negative prompt and "pouting, pouty lips, sad expression" to my positive prompt to get that classic "Traci Pout". Good luck with it!