BERRYBRITE IL

I finally got this model converted over and trained to a point I'm happy with (at least for an initial version). I deleted the previous dataset for the Pony and XL versions a while ago and had to rebuild it, so I decided to make some changes and additions.

To begin, this model was made with the intention of capturing the styles (not the specific characters) of the original Rainbow Brite and Strawberry Shortcake art. And while it does a decent job of that, due to the limited and low quality images available for the data set, the styes have blended somewhat and won't be exactly like the originals.

I also included other art styles to help add more detail and variety to the dataset, which I'll go over below.

Version notes:

V1 Already did a good job of capturing the styles, so I was happy with it for the most part.

V2 Doesn't make any major changes to the overall styles, but it does bring the 'Tmix' style a little closer to the original and fixes the hands in the 'Stsc' style.

I stopped the training for v2 here because I was worried about possible overtraining, since the outputs weren't changing much, and it is pretty faithful to the styles already. However, I may work on one more version, using generations from these models to expand the dataset more.

Illustrious Changes:

First off we have new style tags, some are similar to the tags used in previous models, but to avoid overlap with anything currently in the Illustrious base models I changed the tags to simple four letter combos.

Starting off we have the 'Rbrt' tag (left image in each set), used for the Rainbow Brite style, as well as a new tag for a more detailed Rainbow Brite inspired style based on superhero characters, 'Sprh' (right image in each set).

v1

v2

While this second style was used in training the previous models, I didn't specifically tag it beyond a generic 'Superhero character' tag. However that led to more style bleed in the Pony and XL Loras.

Now it seems to be better balanced.

Following that, we have the new 'Stsc' tag for the original Strawberry Shortcake style. This is the one that took me the longest to get right, due to the poor quality data set, but I was able to fix up the images and get better training results after a few tries.

v1

v2

With this one, the only major note is that another tag is needed to get close to the original Strawberry Shortcake style, 'no fingers'. I know, this seems weird, but the original character all had a more ragdoll-like appearance, and their hands didn't have fingers but were instead mitten shaped. If you don't use the 'no fingers' tag, you will most likely get the versions with fingers, which also come with other slight differences.

In v2 the 'no fingers' tag is no longer needed, as it will default to the fingerless hands from the original style, however to counteract this if you do want fingers you will need to add 'fingers' to the positive AND 'no fingers' to the negative, or possibly include a hand fix/detailing lora.

Next we have the biggest overall change, the former Tkid tags and style were replaced with the newer 'Tmix' tag, and a newer style (though I plan to get it closer to the older style in later updates).

v1

v2As you can see, instead of the pixar-cartoon-esque style from before, this style has more natural proportions. The left side images for each set also include a 'kid' tag, which pulls it back to the older style slightly, but doesn't change it completely. The right hand images for each set don't have extra tags, and default to a teenage seeming character with a general westernized anime/cartoon style.

In v2 it's still not exactly like the tkid style from the pony/xl versions, but it's a little closer, with characters naturally have more child-like proportions. Now including the 'kid' tag (still seen on the left) has a less drastic effect.

Last is the completely new style addition, 'Mnct', which just stands for Mini-cute, as the images used were almost-but-not-quite chibi styles, so miniature and cute.

v1

v2

There's not much to note with this style, I just liked it and wanted to include it rather than make an entirely separate lora for it.

Final notes.

As with the Pony and XL versions, due to the data set leaning more towards girl characters the lora will push more feminine features, so to get males you will have to be more specific with prompts.

Using the 'traditional media' tag helps to better capture the hand-drawn, colored pencil styles of the original images, and just seems to make better quality outputs.

This model was trained using the Illustrious-XL v1.0 checkpoint, but all generations are made using WAI-NSFW-illustrious-SDXL v12 and Fooocus.

All of my models may be used for your own training and/or merges.