This Lora basically adds sex between women and various monsters, demons, and aliens to Flux. The list of trigger tags I've provided is by no means exhaustive, it's just the ones I feel are most important in terms of direct prompt composition.
This Lora works best if you use the relevant tag words within the context of fully formed sentences. You can refer to my sample images to get some ideas on possible ways the Lora can be prompted.
Recommended strength as of V2 is around 0.6 to 0.8, but feel free to experiment as you see fit. (1.0 is almost always clearly too high though per my testing, and 0.5 while sometimes ok tends to be not quite enough).
All three standard XL / Flux resolutions (what the CivitAI generator calls square, landscape, and portrait format) are directly supported by the Lora. Generating locally at even higher initial resolutions may work but is likely to be less reliable (I'd probably recommend generating at a standard resolution and then doing a "hi res fix" upscale from there instead).
This Lora was trained onsite at Rank 16 / native 1024px, on a dataset of images all 1.5x higher resolution than their intended output "bucket" (e.g. 1536x1536 -> 1024x1024, 1856x1280 -> 1216x832, 1280x1856 -> 832x1216).
Description
~128 more images added to the dataset. Adjusted captioning approach on all images to improve separation of concepts and accuracy of outputs. Trained a bit more thoroughly than V1 (40 epochs instead of 35). This version is much more reliable at strength 0.6 or even 0.5 in some cases than V1 was, though I still recommend 0.7 as a good "starting point".
Both V1 and V2 were trained onsite at Rank 16 / native 1024px. All dataset images were ~1.5x higher resolution to begin with than their intended output "bucket" resolution (that is, 1536x1536 for 1024x1024, 1856x1280 for 1216x832, and 1280x1856 for 832x1216). All three standard XL / Flux aspect ratios (what the CivitAI generator calls square, landscape, and portrait) are directly supported by the Lora.
FAQ
Comments (19)
V2 has a small amount of "werewolf" data but not enough that I figured I should put it in the tags as it's gonna be less reliable than everything else. Feel free to try it though.
Maybe you could also make a non-XXX LORA with just the creatures?
I don't think it's that necessary TBH, part of why this Lora works is that my data is able to blend to an extent with base Fluxes pre-existing idea of what these creatures looked like, their physical appearances aren't 100% introduced by the Lora (which is also the case for the women, and exactly what I wanted, I did my best to not impose a "style" that was significantly different looking from a base Flux photorealistic image).
I also don't have a dataset of these creatures that depicts them by themselves without clearly visible penises lmao, so that'd be a bit of an issue also.
Edit: That said I strongly suspect that prompting for what you want using this Lora but at like lower-than-normal strengths (starting around 0.3 to 0.4 maybe) would probably improve the appearance of Base Flux creatures in a non-XXX manner as it is currently anyways.
@mech4nimal yeah technically you can already do this, see here: https://civitai.com/images/28360595
@diffusionfanatic1173 thank you for the clarification. yeah the clearly visible penises where my concern haha now you did it, but with trousers ^^ gonna check it out for myself further. btw: I like your philosophy about using existing content in Flux without bending it too much, because some LORAS that try this may achieve a goal with certain subjects, but also seem to introduce new problems on the way.
@mech4nimal No problem! And yeah after testing more it still inherently wants to put pants on them basically at lower-than-normal Lora strengths lol (and it becomes much easier to add stuff that wasn't in the dataset, like the silly cafe ones I did and the picnic one if you look at the crosspost onsite generator gallery now). Around 0.4 - 0.5 Lora strength seems good for still applying the visual aspect of my creatures to what Flux thinks those words mean but in contexts that aren't necessarily NSFW. (In some cases you can still do SFW even at like 0.7 though but I'd still probably start lower than normal if you're going for SFW).
Something I should note: the Lora strength was not actually 1.0 on any of the sample images, nor was the "CFG Scale" actually 1.0, they only come up like that when "remixing" them because CivitAI can't read ComfyUI metadata properly. There is nothing I can do about this, unfortunately. Change the "CFG Scale" to normal default 3.5 and put the Lora strength to 0.6 - 0.7ish if you are trying to directly remix any of those images.
V3 plans
- maybe expand dataset a bit more
- swap out some existing images I'm not completely happy with
I've noticed that often the frontal legs of the dominant part are not on the side but rather behind, kind of weirdly hidden or glitched away. If you get that improved, it would be awesome. I already like the lora a lot.
@lost_moon do you mean for like the more quadrupedal kinda monsters that it'll sometimes do?
@diffusionfanatic1173 Yeah quadrupedal. I usually specify that. It does work pretty well already, just needs a bit luck to get the front leg right. :)
I will be doing an SD 3.5 Medium version of this for sure (already trained actually, testing now). I don't think I'll be doing a Large version however, as to me the better prompt adherence is not really better enough to outweigh the more limited resolution and generally subjectively worse image quality of Large.
This is still happening BTW, I just have many different projects I'm working on all at the same time lol
Hi ZootAllures9111, I have read quite a few of your comments around reddit and you seem to be one of the sensible voice about how to properly train a lora on flux. I have trained a dozen of them so far with mitigated success. I am trying to nail a very specific character lora but none of them look right. I use mainly kohya_ss scripts using fluxgym gui. Can you summarize what you mean by "captioning in natural language at first but followed by tags for accuracy" ?? Is Dim16 enough for a character Lora to remain flexible yet accurate? What learning rate and settings would you use for 20 images in the imageset, using flux dev? I would love to pick your brain if you'd be willing to chat directly on discord or something. Thanks!!!
This would be so great for Wan ^^
Male penis in Wan is difficult, but adding anything other than a man makes it really difficult to impossible.
That would be great, there's nothing about that currently, nothing at all.
I mean there were no human males in the dataset for this whatsoever, I'm not 100% sure what you mean lol
@ZootAllures9111 I didn't ask for male data set.
If you could do this for Wan 2.1 or 2.2 too that would be great
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.