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
Initial version. 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 (9)
For V2 (likely to happen, with an expanded dataset) I'm thinking probably ~5 more epochs than this version, should bring the overall "hit rate" up even more.
In case anyone is wondering, no, you absolutely cannot make a flexible Flux Lora like this with "no captions" or "just one trigger word", it's completely impossible for the exact same reasons it always was in SD 1.5 and SDXL.
I'm laughing at how calm the women look in the renders!
Lol yeah, I might try to include more expressive faces in the dataset for any V2 I do. I was moreso focused on trying to include a diverse enough selection of ladies that it wouldn't impose any particular "look" too much, for this version.
Edit: that said you can probably "prompt in" specific facial expressions to some extent even currently, I never actually tried this while testing the Lora TBH.
Just started training on V2. ~100 more images, and slightly adjusted captioning approach. Doing a non-Pony SDXL version simultaneously also on the same dataset cause I figure it'll work just as well and some people might want it.
To follow up: yeah, you can just prompt for specific facial expressions on the lady if you want with this as-is already, works fine. A couple images I've made onsite in the last little bit do this as an example.
Fun! :D
How on earth did you manage the positions!!!????
Edit to add that you can definitely change the lady as well ;)
I just trained it the same way I'd train any Lora for the most part, no real difference in captioning approach beyond putting some key phrases up front given that "caption shuffle" is not available as a training feature for Flux. V1 is ~415 images, trained with Flip Augmentation turned on. V2 (out soonish) adds about another 100 images on top of that, and slightly adjusts some aspects of the captioning.
I basically feel that people are massively overthinking Flux training for no particular reason, accurately describing as many elements of every single image as is possible is still the overall best way to go for this kind of brand-new concept Lora on Flux, as it always was with SD 1.5 and SDXL.
The only reason you wouldn't be able to change the lady is if you used an overly-minimal tagging approach (and that's not different than on SD 1.5 and SDXL, captioning that way was equally bad on those models for the exact same reasons).
Testing out Flux V2 and the SDXL version currently, both have actually a better overall "hit rate" than Flux V1, I think due to the slightly revised captioning approach I used.