Update: New version for SDXL, retrained with some of the results of the first version in order to improve style flexibility and some other issues. Should find it better than v1 in most situations, but it may be slightly less accurate on character details.New LoRA for the monsters from "Rumble"! This one mainly is intended for Illustrious/Noob models, I wasn't able to get it working well on Pony or SeaArt as I usually try to. May attempt that again in a later version once I have some good synthetic data. Style flexibility is decent, depending on the finetune used to generate. Some characters have a few oddities (see notes below), and there are some biases I wasn't able to fully separate from the model, but they can be managed with negative prompting/negpip or some tweaking of the LoRA weight and refiner stages. Overall just keep in mind this is not a "good" LoRA at this point but I mostly just wanted to see what I could do with these characters.Here's all the characters included and the suggested tags to improve consistency, describing other physical features may also help:axehammer (female, iguana, yellow scales)<ul><li>her nose horn and spike tail can sometimes end up in odd places or shapes at higher LoRA strength</li></ul>denise (purple and white fur)<ul><li>since this character is just kinda a walking ball it can look strange making them "anthro" though it does work</li></ul>king gorge (male, bulldog, horns, brown fur)<ul><li>often generates with a crown but that can be removed with neg prompt</li></ul>klonk (male, boar, blue fur, purple mohawk)<ul><li>fur color is somewhat inconsistent, if not specified will often end up tan or orange</li></ul>lady mayhem (female, toucan, blue feathers)<ul><li>can give her white/tan countershading sometimes</li></ul>lucho luchon (male, dragon, blue scales, white countershading)<ul><li>his luchador mask is pretty much baked into his design, could try removing that with neg prompting but not sure what it would really do</li><li>specify "taur" or similar to get his normal body type, otherwise "anthro"</li></ul>mcginty (male, aquatic, anglerfish, green skin)<ul><li>often ends up generating with headphones as most training images include those, but can be neg prompted out</li></ul>nerdle (male, 4 eyes, yellow fur)<ul><li>his outer eyes don't always behave as expected, ex. if you prompt for "eyes closed" only the ones in the center may be. Can also end up looking more like black+white cheek spots in some cases</li></ul>ramarilla (male, gorilla, horns, grey fur)<ul><li>sometimes seems to end up with a tuft of hair like a mohawk</li></ul>rayburn sr (male, horns, yellow scales)<ul><li>all of the movie images of him were sepia-tone flashbacks so I tried to color-correct the training images, it may have made him more yellow than intended</li><li>will probably also just work if you prompt for "rayburn"</li></ul>steve (male, orange body, horns)tentacular (male, shark, blue skin)<ul><li>doesn't fully pick up his odd tentacle-arms in most cases, there may be some way to prompt for that more consistently but I didn't look into it much</li></ul>ting tang (canid, brown fur)<ul><li>technically this is two characters, the "ting tang twins", but since they appear nearly identical (and I don't know which is ting or tang) I combined them into one</li><li>like lucho luchon, use "taur" if you want 4 legs and 2 arms, otherwise can do "anthro" easily enough</li></ul>yokozuner (male, lizard, red and white scales)<ul><li>sometimes seems to overlap with axehammer a bit, and does some odd things with his "topknot"</li></ul>I didn't try training a specific trigger for the style in this set, as there wasn't much distinct about it beyond being a 3d-animated movie. Most of the movie images were tagged with "3d animation" or something similar, you can try playing with the weight of that to see what effect it has. May improve character "accuracy" at the cost of style flexibility. Also images will often end up with wrestling ring-style backgrounds, as that's common to most of the training images, so negative prompt/negpip that if it keeps showing up. In general there can also be some difficulty in getting clothes to appear correctly, as none of the training images show characters wearing clothes. Hopefully these issues will be reduced once I can get a good set of synthetic images from this version and reincorporate them into the next training run, but will try training some of the other datasets I have first.

Rumble - Multi Character LoRA

Updated version for SDXL models, incorporating some of the outputs of the previous version. Should have better style flexibility, less issues with clothing

Files with SHA256: 7d8bc1db7b1e97aef46001857276f1742a5caffb9b6cba07185c615581b332e4