American glamour and pin-up model, the original one. There is no model for her on Civitai yet, though I'm aware she is embedded in the base SD 1.5 (that version feels a bit different though). You can also combine them for an interesting effect. The LoRA produces both color and monochrome images, you should specify it in the prompt if you want just one of these. Added a TI version, using LoRA with TI dramatically increases resemblance.
Description
FAQ
Comments (12)
Prompts and modals?
Prompts are simple, something like "monochrome photo from 1960's, woman with pitch black hair in photo studio". Checkpoint is AbsoluteReality 1.6. Images made with ComfyUI, so full workflow is included (just use Copy Nodes option).
Doesn't have the likeness at all, barely has the haircut.
Thanks for the feedback, I think the likeness is also related to age and use of make-up, you can adjust both in the prompts yourself. You can also try to prompt for a specific haircut, it should be pretty flexible for such adjustments. In the worst case, you can keep using the base SD 1.5 version if you prefer, this is just another option.
@dolirama126 Keep on working on it I don't think anyone has really nailed her likness yet, atleast that I've seen. Some faces are more challenging than others, Margot Robbie is another SD has trouble with.
Once you or someone really gets it perfect though I'll be the #1 fan, would love to make Bettie Page Pinups of my own design.
v2 is a bit closer to looking like her. The hair and face are both better. SD seems to have a problem with the shape of her face in general though. In some gens using your v2, it comes out looking pretty close. In other gens, the face is too round and doesn't look like her at all. Judging by your samples, it looks like you encountered the same issue. Hoping for a v3.
Round face is a result of AI's RNG - you simply can't and won't get the correct face all the time regardless of how well your model is trained. I generated about 40 samples with this version, half of them was pretty good but can't post them because of ToS. Half of the other half is what I posted here, and the rest had some problems (not always related to the face), but that's to be expected. So these are not cherry picked results, this is pretty much an average output of the model. And if I can get good resemblance every other generation (in average), I consider it a success so v3 is unlikely (at least not in the near future).
@dolirama126 You will get a CORRECT Face if you start using a "Trigger Word" to
add into your training images. I notice you hardly use Trigger Words.
Example , choose a word which your A.I. will not
understand or there is no other images with the name "Bettie Page" or "Betty Page" in
the Stable Diffusion A.I. main program.
Choose a word which in it you also exchange some letters for numbers.
for example - an "1" will be a "1", an "e" will become a "3".
-
I suggest to use the word "b3tt1epag3" as your trigger word and this word MUST
BE USED in the training of the Lora Model.
-
When you train the model in the Stable Diffusion, when adding a "trigger word"
the A.I,. will work better when using the LORA or Textual Inversion, otherwise
you will spend too much time creating images which will not look exactly like the person
in the original pictures you used to train the model with.
-
If you keep creating models without trigger words, you are just spending your time
on making models which will be more difficult to use for you and everyone else.
I KNOW how long it takes to make a Lora model (1 to 2 hours), so time is VERY important
and you need to use a trigger word for everything to be created perfectly.
Looking much better! You nearly have it.
Bettie has this hot thing she does where she kinda scowls and looks angry/mad. It makes her extra hot. Either that or the happiest smile in the world.
Your samples here have her more neutral, so it looks like her face, but not her character.
I love it though, keep up the good work!
Thanks, that's useful feedback. I think I have used mostly her "angry" pics for this version to make it consistent, because what you refer to as "happy" pictures actually also make her head shape change to round-ish. So mixing these two types of faces in the training dataset is what caused the first version to be somewhere in the middle and in the end not looking quite like either of both Betties. The average output you get from this version (basically all the samples here, I didn't do much cherry-picking) is the angry look, and that's without using any emotion-related prompts - I have not tested that, but you could try it for better effect.
@dolirama126 hmmmm maybe she could use a n angry version only, a happy version, a neutral version, and then try to do a merge.
@AstralNemesis She would probably need at least 2 versions you mentioned if you want a perfect model indeed. But you can try to combine these two LoRAs plus the "Bettie Page" SD 1.5 knows and play with their weights - it could give you even more options and perhaps one of them will be satisfactory.