Based off of early '90s Jim Lee's art style with modern colors and inks as presented in Mutant Genesis 2.0.
More Versatile Version
More stable. No need for negative prompts below.
Keep weight under 0.8 and if you're having difficulty feel free to drop it to 0.5.Works better with Pony Diffusion than the old version, but best results still arguably with MeichiDarkMix_Reload and WAI-ANI-NSFW-PONYXL.
The trade off for more stability and cleaner visuals is that it is often less "Jim Lee"-like. So if you're looking for a more authentic Jim Lee look go with version "V1" and drop the weight to 0.5-0.6 and use MeichiDarkMix_Reload.
V1
So far best results for checkpoint are with MeichiDarkMix_Reload and WAI-ANI-NSFW-PONYXL. Pony Diffusion and Pony FaeTality also work, but not as well. Weight at 0.8 or less has led to best results (0.5 if you're looking for no blur). It's probably best to remix from one of the samples below but you do you.
Prompt: j1m l33,score_9, score_8_up, score_8,
Negative prompt: irregular border,word balloon, speech bubble, english text
Description
FAQ
Comments (9)
Requires some work to get working, but works if you do.
My recommendations:
1: 42 images (according to metadata) is short for a style. The number needed for a style isn't as well documented as for characters (which is itself not well documented), but I'm fairly sure it's at least 100 unless the style is only for something of very narrow scope (like only portraits, or redraws of a particular meme).
2: For comic data sets, I recommend blanking out speech bubbles and using the tag "blank speech bubble" (alongside the base speech bubble) and skipping any issues with text entirely.
Thanks for the feedback! I see mixed opinions on number of images for style. Although I gathered about 140 initially I ended up just going with the 42 highest quality images to see how it went. I would like to try again with a larger dataset just to compare results. I see people doing Version 2s and up all the time but I don't understand how to do that. Can you point me to a resource that can explain this?
Also, I trim the images heavily which probably isn't as important for style as it is for characters, but I mostly do it to avoid unintended panel creation and problems with text. However, it causes the "irregular border" problem in images set at high weights. Any idea on how to avoid that issue?
@superneurotypical I tag images with visible panel dividers as "comic", since that's what the auto-tagger suggested and it's standard tag for multi-panel images on boorus. Haven't had an issue, though such borders are the minority in my training data. Beyond that and "speech bubble, blank speech bubble", I also add "sound effects" where applicable. For a style, I also want to add that you shouldn't neglect tagging the backgrounds/background elements where possible (auto-label is normally decent here). Edit: Looking at gens and metadata some more, you also want to use "muscular" where applicable in training data, and include some more pictures of men and (if possible) backgrounds without characters (versatility in subject is important to style LoRAs, even if you, understandably, just want to generate pics of hot babes).
As for V2s: It's really just training the LoRA again from scratch with changes to the data. Only complicated thing is that, if you're using the on-site trainer, you have to download the result and upload it to have it as as a version. If you're asking about the actual uploading as a V2, it's started by using the "+" button to the left of "V1" in the upper left of the page (below the name, updated date, and tags)
@NanashiAnon Thank you for taking the time to explain this! I've been wondering how everyone else managed to avoid the problems I had with this and other style loras I've made. Excited to get moving on V2 now.
Awesome style!
It's super blurry at high weights and outside of MeichiDarkMix_Reload, but it's the best I can do!
If you're having issues with site instability, I recommend scheduling your posts (click the clock to the right of the "publish" button when you're asked to make the initial gallery). Gives it plenty of time to process everything, and it will still appear under "newest" models. (I personally set mine to publish at midnight GMT because that's when buzz rewards for reactions reset, but just few hours should be enough). You can even generate with it to confirm functionality/make an initial gallery under some conditions (I'm not quite sure what they are, but direct import from on-site generator definitely works. Posting as v2 to a model that was trained on site also seems to work).
As for the older pictures needed for a style thing I mentioned for V1. From my further research I think the number may be a bit lower than 100. I've since gotten OK results from an experiment with only 56 (goes up start of Monday GMT), but that is really hit or miss on where and how it applies the style.
Yeah, site instability has been a real pain. And when I upload it as a version upgrade then it defaults to Dreamshaper and will not work with any Pony checkpoints. It doesn't matter that I set it as Pony and used the same data and it generated just fine with Pony when I uploaded it as a new model. Maybe it just needs more time to cook. I will try the time delay thing if it happens again.
I checked with ChatGPT a couple times on sample size. It said a few different things, but generally said 100 at a minimum for Style Pony is a good idea, though over 200 would be better. I know I solved some of the issues in this version (if I can ever get it to properly publish...) but there's still some facial imperfections at higher weight that bother me (it's also still picky about checkpoints, but less so). I think increasing sample size will help though I also need to figure out exactly what to do with num repeats and epochs.
@superneurotypical I think the defaulting to Dreamshaper is actually the site just not having the LoRA into the system yet. You'll notice it doesn't actually load the LoRA. I've seen it happen on other new LoRAs over the past few days and it cleared up within a few hours (hence why I recommend scheduling)