Making all the characters from a series with screencaps alone? Why not add a style?

To make my Jimmy Neutron Style LoRA, I went and made LoRAs for a bunch of characters in the series to get tagged training data for the series before combining them. I see other creators do 6+ characters from the same series with screencaps only often enough that I wanted to encourage people to do it the other way around (combining their existing datasets into a style) because well made styles for particular works are both a popular yet under-served (because of just how many tagged and selected images these require to be good) type of release.

Part 1: Prepare a supplemental package

First, get screencaps with no people in them (objects, scenery) the same way you got screenshots of the characters you made. If your dataset consists only of females, get some male characters. Characters who have a distinct body features you can narrow down to some existing tags ("muscular") are potential supplemental source. Women in underwear/swimsuits/nothing that wasn’t in your existing datasets is also worth grabbing if it exists since it should help make the final thing better at scantily (if at all) clad women. Crowd scenes are also good for character variety. 40-50 images is more than enough, this just to give it more to compare than just pretty girls.

Part 2: Prepare the merged dataset

1. Make a new folder

2. Make new folders in it for each dataset you’ve completed

3. Extract your dataset to those folders

4. Delete the images+tagset you used that were redundant for non-characters learning (e.g., same background) from these copies

5. Zip up the containing folder (don’t worry that the names would overlap if not nested)

Part 3: Tag

This will vary based on your trainer, but for on-site trainer

1. Upload and tag the supplemental images. Since you’re not trying to get tags totally consistent here like with a character/concept, you can get good use out of autotag and really only need to correct major errors (1boy vs. 1girl)

2. Upload your merged dataset

3. (Optional) delete the triggers from your tag list (Since you axed so many images they likely won’t work well anyways)

4. Add a main trigger for the new thing

5. Train as you normally do. If you want higher quality in exchange for more buzz, you can up the repeats (note that above a certain threshold of total images these are automatically dropped to a lower value by default)

6. Publish it so everyone can enjoy it