This LoRA have eight Simpsons characters! Generate images of Bart, Homer, Marge, Lisa, Ned, Milhouse, Luann, and Mary
V2.0 seems better - less steps better results.
Triggers: bart_simpsons, ned_simpsons, homer_simpsons, lisa_simpsons, luann_simpsons, marge_simpsons, milhouse_simpsons, mary_simpsons
Character-Specific Tags:
bart_simpsons, orange shirt, blue shorts, yellow skin, spiked hair,
ned_simpsons, brown hair, mustache, glasses, shirt, sweater, yellow skin
homer_simpsons, white shirt, homer mouth, bald, yellow skin, pants,
lisa_simpsons, orange dress, pearl necklace, spiky starfish hair style, yellow skin,
luann_simpsons, blue short hair, glasses, earrings, white shirt, purple pants, yellow skin,
marge_simpsons, blue beehive hair, necklace, dress, yellow skin,
milhouse_simpsons, glasses, straight blue hair, shirt, red shorts, pink shirt, yellow skin
mary_simpsons, orange hair, short twintails, freckles, blue dress, yellow skin,
Description
v1.0
FAQ
Comments (11)
In my experience, the more characters you put into a Lora the less stable it becomes. But this one seems to be holding it together well enough.
Thanks for the feedback! I really appreciate you taking the time to test it out. You're right, multi-character LoRAs can be tricky. This is actually my first attempt at one, so I'm still learning the ropes. I'm glad to hear it's holding up well so far. I'm definitely planning to iterate on it and improve it as I learn more.
@terpentinas I find the opposite. Maybe because it's trained on WAI as a base and not the default Illustrious, checkpoints that normally are fine with loras like this are creating some real body horror instead.
@karetai Thanks for the feedback! You're right, I trained this Lora on WAI-NSFW-illustrious-SDXL, because it is the most downloaded base Illustrious model on Civitai. Same Lora works differently on different base models. Some base models are more similar to each other and results from Lora are more similar. I do not know how to solve this issue (would like to find a way especially for realistic Illustrious Loras). If someone knows how please share.
Only solution I could offer right now is to train on another Illustrious base model.
@terpentinas It's usually a good idea, unless there's a reason not to (e.g maybe some realistic subjects don't train too well on Illustrious), that one should train an Illustrious lora against base Illustrious. It'll be the most widely compatible within this variant of SDXL. If you train specifically on WAI, for example, effectively that means this Lora will be far more tied to the concepts and styles that WAI has added. It might be worth testing a version trained against base Illustrious to see if it's more broadly compatible with other Illustrious-derived checkpoints, if you cared to do so. If you do use this current version with eg. knk Luminai, or Coco Illustrious, it's like an episode of Treehouse of Horrors!
@karetai One of my galleries is what happened when I tried a Bart Simpson Lora meant for SDXL in Pony. Those were literally the first images I generated and also one of my favourite things I have ever uploaded to the site. So I think there is some value in a Lora acting weird in certain circumstances.
@tombot If a Lora is trained against base SDXL, it likely won't work (or work well) with Pony, or Illustrious, etc. Think of the foundational SDXL model as a "brain" of fixed size. In order for these large finetunes like Pony, Illustrious, Animagine etc. to work, they have to "swap out" a lot of the information in the base SDXL model, as you can't really (effectively) increase the size of the data in SDXL. That "loss" of other information is why the Loras are largely not compatible, if they used any of the paths that have been trained over. You'd need a whole new foundational model if you wanted more data or potential, such as is the case with Flux, SD3/3.5 etc.
@karetai I know this lora need improvement, I want to improve it, I tried your suggestion and trained it on illustriousXL_v01, was alright, still had many same problems, that I can fix later (need a little time because it is trained on 441 image, takes a lot time to correct descriptions and images, but I already recognized some of my mistakes like, homer has no hair trait and many more, to find way to improve interactions between characters so fixing this problem will not be that fast like just training it on a different model (that I still willing to do). There is some problem in my training that one random seed gives expected results, and other really bad ones. I tried knkLuminai_v10 and cocoIllustriousNoobai_v56 and it seems ok for me. Shared a post with a simple test. Could you share yours? Maybe there are more problems I don't see, I really appreciate your feedback. It helps to improve.
@terpentinas I'm only going on my beginner knowledge, and I appreciate all the work that others do, so thank you for experimenting and trying things! What I've found is that tagging plays a big part, not just what's tagged but also what's not. From my own limited experience, let's say you tag "milhouse" in addition to "glasses". It'll "learn" that glasses is a separate concept to Milhouse. And so to have Milhouse with glasses, you'll need to specify both - and then it'll usually try to put glasses on everyone else in the picture, too! It's a lot of experimenting, even my own non-published loras I find each one has its own unique challenges or things it doesn't do "right".
In my experience its perfectly achievable. Just a matter of good prompting. Make sure the characters have proper different descriptions and adding a/b-tests into the data. Literally adding example pictures with the same image 3x, but erasing one character + their prompt, so AI learns what you mean.
WOW!!