Last updated - 21/03/2026
I now have a Ko-fi for those that'd like to support my projects. While nothing there will remain exclusive, you'll get more substantial progress updates each month, and access to my own non-AI images at higher tiers.
Model description:
This is an attempt at an all-in-one lora for generating the monster girls of MGE, which allows for extensive customisation and feature-mixing. Refer to these articles for more information. Not everything works perfectly, so good results may require a bit of experimentation at times.
For updates on future versions, I generally post what I'm working on in my profile's announcement tab. It'll probably be MGE-related, but in addition to this one I occasionally make standalone concept-loras for monsters that would benefit from having one.
Background info:
"MGE" Refers to the monster girl encyclopedia book series by Kenkou Cross. For more information, see the MGE wiki here.
Additionally, while I'm not affiliated with them, the wiki recently opened donations so you can donate here if you'd like. They have a $12/month upkeep cost apparently, so I'm sure they'd appreciate the generosity. Be sure to mention where you got the link from too, for the model's sake :).
"Non-entertainment" usage:
Any "commercial projects" must be transformative, so no posting your results to Patreon and asking for money. Using it as part of a larger project is fine though, and I'm happy to help if you reach out with questions, or simply to tell me an interesting use-case you have. Bear in mind that if you're creating stuff linked to MGE specifically, they must also adhere to the creator of MGE's terms here.
Actual artists:
If any artists dislike the idea of their stuff being used for this (and you've drawn things from MGE), let me know and I'll avoid your work. I haven't done this preemptively due to the huge number of dead source links in the archive, but if you reach out (with some proof of identity to avoid trolls) I'll remove any of your content from my dataset/s.
Reuploads:
If you reupload this, please link both this model page and my taglist for ease of use. I'd also appreciate you letting me know where you've uploaded it, so I can keep track of where it's being used.
Description
FAQ
Comments (30)
Can we get a list of all the monsters that work?
There's now a list of everything I've tested in the main description. I haven't been able to test everything yet, but generally the more species that share a certain feature, the better said features turn out.
Tempting, but the file size really puts me off.
I don't believe there's much I can do about that without reducing the quality, as there's ~500 images and I used dim/alpha at 128/32. V1-V4 are smaller, between 110mb and 550mb iirc, so I'll upload them here if there's a demand, but they're generally worse than this one because I barely knew what I was doing when making them, and the tagging is a lot messier, mostly just autotagged.
(For a general idea of the quality difference, the dragon-zombie preview image was from V2, the raiju/slime carrier and dragon character sheets were from V3, and the succubus was from V4.)
@Blizzard264 Yea thing is though LoRa models are not usually that large in size. That's gonna really impact rendering times. Plus properly trained LoRas will still output really good quality since they don't override the main checkpoint, they assist it like an overlay filter (to put it simply) 1gb is way too big!
The fact that no one has posted images of using it should tell you that it doesn't produce good results. I'm sure in your testing it works fine (which is what every model maker says that gets offended) but you need to remember that everyone has different PC builds and having a LoRa 1gb is not gonna make things run faster or look better. I've tested this one on my build with my RTX 4070 and it is slow as heck and at a weight of 1 is does not produce good results. most images are overbaked and poor quality. Running it at a weight of 0.6 doesn't even trigger the LoRa very much, so theirs no point to using it.
Maybe change your approach on file size and data sets? Larger doesn't always mean better. Optimized is always going to win with better results.
@LindezaBlue No offense taken, I appreciate the feedback, I'm quite new to making these and I haven't been able to find much (in depth) information. From what I could gather, increasing dim/alpha size "helps it store more information", which I assumed would help more in a lora like this one, where it's a couple hundred "characters" as opposed to one specific person, style or feature.
My first four versions used 32, 64 (for both V2 and V3) and 128 dim respectively, and V5 is 160, but if the file size is an unnecessary thing I guess the quality increase must be due to other changes I made rather than "bigger=better". I'll look into the effects of reducing size for the next version (which'll take ~a week to get enough buzz for a proper train).
The dataset is about as optimised as I can really get it though, besides a few bits of fanart I could probably cut out. As it's trying to cover every monster girl in MGE, most ot them only have 1 or 2 images, while the rest is trying to cover the more obscure features that I've had issues reproducing.
(Btw, there's a bunch of images on other versions, I posted this one today :)
@Ginji23 @LindezaBlue If you're still interested, V6 is only ~700 mb. I'll look into the effects of making it even smaller with V6.2 onwards.
It would be super helpful to get a copy-pasteable tag set for a stable generation of each monster girl (for example, what combo to use for a good harpy or bunyip). I know you lay some of this out in your notes for us to figure out, but it's more an accessibility convenience.
(I know I can also go off your example images, but those have other noise in them rather than base tags for each girl).
Noted, I'll see what I can do. It'll probably take a while, but when I get a bit more free time I'll focus on the more popular species first, and work my way through.
@Blizzard264 Yeah, I tried a few things and couldn't get it to really work right. I think each of the girls seems to need a lot of helper tags?
Some do, some don't, I think it's just down to how many images are already in other model datasets under the same name. Lamias, for instance, tend to just work, because there's a lot of them throughout media, and they all look pretty much the same, but something like, say... the manticore, which is completely different to the "manticore" images most models are likely to have in their datasets, would need help.
@Sargovel I've started a template section at the bottom of the main description. I'll try to refrain from having a bunch of duplicates, so anything with animal limbs uses the beastman template + its own thing (eg: raiju would be "beastman template", followed by "raiju template"), while reptilian monsters like dragons are under the reptilian template, and so on.
@Blizzard264 You can see i've been playing around with this. Much better consistency on my end and me pushing it a little with lora mixing.
Some trouble with stuff like manticore and bunyip, of course.
Yeah, it does tend to work quite well with other loras. In hindsight, I was probably too specific in my tagging, but once you get a stable prompt it makes it quite easy to add/remove specific features without breaking things.
The manticore is to be expected, I've had a lot of trouble getting the tail tip to (barely) work myself. Not sure about the bunyip though, you could try specifying "lamia tail" top/underside, or something like "scaled lamia tail" in negatives? It also could just be a case of the step count though, as the previews were all done on Yodayo using 100 steps.
Oh, also, I'm not sure whether it's a model thing, but using "full body" tended to decrease the quality of the actual girl, despite not usually changing much. It could be a case of the stuff I've tagged with it in training coincidentally being lower quality too, so I'll look into that. If that ends up being the reason why, then using "(thing) fully visible" or "(thing) out of frame/cutoff/etc." in negatives might perform better for full-body shots.
@Blizzard264 Tried some of this and also playing with 5.2.
Regarding the bunyip or lamia tails in general, have you figured out any way to consistently generate a coherent/contiguous tail?
@Sargovel I've edited the Bunyip template slightly, so it's now "furry tail topside, scaled tail underside", with colors separate for both. I'm still having issues with "full body" quality in general, but putting things like "close up, tail offscreen" and similar stuff tends to help if it keeps being cut off.
With getting a single consistent tail though, aside from using "consistent features" (and/or "inconsistent features" and "random tail positions" in negatives), the best results I've had are when there's not much detail elsewhere. (So white/black backgrounds, and fairly simple character details). "Side view" might also help? I haven't tested it with lamias, but it might get a better view of the whole thing.
@Sargovel Out of curiosity, have you been generating at a smaller size and upscaling? I've been messing around with image sizes, and I've found that things seem way more consistent when I'm generating natively at or above 768×1024.
@Blizzard264 I basically never generate at a smaller size and upscale - that's with the extra tab, right? I'm not familiar with doing it really.
I typically just set a 512 x 768 and then either 1.5 or 2x Hi Res Fix
@Sargovel that's what I meant, sorry. But yeah, that might be part of your issue, as in my experience it's harder to get fur to generate consistently at that res. (Emphasis on "might" though, it might also just be my prompts)
@Blizzard264 So what do you do for upscaling after generating your native one? I can start trying that.
@Sargovel Keep using hires fix, by all means, it's just that the thing before the 1.5/2x seems to give better results when it's larger, particularly for large amounts of fur. (So like, if you wanted a 1024×1536 image, you'd generate a 768×1024 image and do a 1.5x hires fix instead of 512×768 with 2x).
As I mentioned previously, I dont know if this is actually the thing affecting my results, but changing it did see general improvements.
Do you ever intend to add stable prompts for the rest of your list that don't seem to tie into the templates? Like, say, for khepri or mindflayers? I tried a few of those and didn't seem to get anything from my initial testing
@Sargovel Sorry, I didn't get a notification for this one. Yeah, I'll get around to it, I've been tinkering with my local setup (AMD card...) for a few days to get it into a usable state, so hopefully I should be able to test things a lot faster than before.
The mindflayer in particular, though, has been very finicky, as there's very few good pieces of fanart to reference. I'll keep trying to get it working properly, but that one in particular will probably take a little while.
@Blizzard264 No prob; as you may have seen, I basically went through most of the list and tried to kind of mix and match the templates with some of the other keywords in the data set. I somewhat forced it to work for some, less so for others. If there's a semi-reliable set that works for you for the individual options, that'd be sweet.
Admittedly I still do have trouble just doing the basic stuff like bunyips if i deviate from samples (or even if i don't it spawns multiple tails or forked tails), but it's generally fine with the templates I see so far.
Any winning combo on arachnes/spider girls? Mine ain't coming out so hot. Surprisingly, centaur actually works despite not even being on your list, but not arachne.
@Sargovel Someone else asked about arachnes recently, but it's probably a dataset thing, which I'm looking into fixing with the next version. Good to see that centaurs work though, as there's not a great amount of good fanart for them that shows the full body...
I do appreciate when you update the templates by the way, I've noticed you periodically tweaking them.
@Sargovel Np, I'll add the new stuff too once I've done a bit more testing. (Also, ty for the tip)
Given the tangled web of updates and changes between versions, an updated template set for possible girls would be very useful...
@Sargovel I'm working on it, it just takes longer to test things now as there's far more that actually work, and a lot of older ones use new tags. I'm making a purely alphabetical list rather than the wierd template system this time, so there's ~230 to go through.