All girls were 18 or older at the time they were models!!
A long, long, long... long, time ago. People used their phone landline (that's a phone that had to be plugged into the wall for you youngsters) in order to connect to the internet. Your computer would basically make a call, there would be beeps, static, hissing. Your computer would sound like it was screeching at you, and only then could you surf the interwebs. It was really slow, so slow that "HiRes" was a 640x480 video and you would have to actually download it to view it. Then you would find out that you waited 30 minutes to see a 10 second clip that by today's standards was more like an animated GIF. Then you would realize that you got much more satisfaction from images. On a good day when the phone lines were clear you could actually view a whole set of images (about 15-20) in just a few minutes.
That's where these girls come into the "picture". Back in the late 90's and early 2000's these 8 girls were a few of the most popular teens on the internet. There were no cam sites, and no home run membership sites. Just a bunch of small studios trying to jump into the new Millennium by paying college tuition for pretty girls.
The Girls and their tags
Nicole - Nicole Scott, aka Red Hot Lauren
Ariel - Ariel Rebel
Ivana - Ivana Fukalot
KittyJ - Kitty Jung (had to add the J to avoid cat ears all the time)
April - Little April
Melissa - Melissa Ashley aka Anne Howe
Fransizka - Franziska Facella
Tawnee - Tawnee Stone
points of interest
-All girls were of legal age at the time of posing despite what their body type might imply. (I.E. KittyJ and Melissa)
-All dataset images were cropped at a 3:4 ratio. About 1/3 of the images I used to train this LoRA were already in this format, and it is also what I use most of the time. Likely will not matter when rendering but I'm just throwing that out there.
-Each girl was trained on 32 images each. All images were pulled from the same site and still easily accessible to the public.
-The LoRA is named Rev2Final because I almost had it ready to release but tried something new before I uploaded it. The end result turned out just a little better and I had thought about releasing both to allow for options but after some testing Rev2 is more capable.
-Final does not mean that I am not considering updating it. I'm actually in the process now of adding 8 images to each girl focused solely on facial features. I am also thinking of making it an even 10 since I remembered another from that time that I missed.
-Not all girls turned out to be as good as I hoped for. April for example never seems to be quite right and although Tawnee is close enough to be recognizable, she is not good enough for me. Luckily Nicole turned out mostly good or I would not be uploading it yet.
error in training found It would seem that I forgot to include April's training images so I have removed that trigger from the list, 7/8 ain't bad right?
V1.0 through V10 were a learning experience, ignore them.
V2.0 - so I learned a few things. Not all girls work at full accuracy, yet, but I'd say they're all 90% correct with a couple getting 95% or more accuracy.
Description
This is not actually a new version of the Millennium Girls LoRA. This is a test LoRA that has 5 girls similar to the existing 10. I wanted to test a theory that a LoRA could be trained on far less images if they were similar enough to each other, yet still different enough to form a dataset. This "version" was trained on only 50 images, 10 for each girl. The smaller dataset means I was able to custom caption the images. The fact that the images were similar (same set) means that most of the caption files were or were nearly identical to each other. Choosing to add only 5 girls means that it was easier to avoid similar captioning between girls.
Remember This is not a new version but a supplemental LoRA, you will need to download Redux to have all girls.