If you like the model, please add image to the model page + Add Post and leave a review, it really helps. The buzz makes it easier for me to make improvements/train new models.
v3: i messed up and uploaded the wrong file for previous version, this was the intended file. sorry.
with some other loras, it will look wet, if you don't like it, just negative prompt "wet, shiny"
v2: still not perfect but now consistent hair stays inside and matted and doesn't overpower non realism styles, thanks to everyone that donated buzz, I will update mentioning you if you wish, just let me know.
Prompting: tight improvised mask, wrapped head, sheer fabric mask, blindfold, gagged
the closer the style is to realism, the better it looks, but now does transfer to other styles unlike the first one. two more attempts since the last one.
You can lean into the realism by using a realism checkpoint and it looks the best there.
v1: burned a bunch of buzz trying to get this to work, not to my liking but can't figure out how to make it work without overtraining it. might as well try to recoup some of the buzz cost and it can work with very specific prompting but struggles with some styles. maybe i will try again
The insex mask is pantyhose pulled over head with tape blindfold and gag over it.
Prompting: ins3x mask, improvised mask, sheer fabric over head, encased head, blindfold, tape gag, flattened hair, face compression, tight mask
Negative: ((loose hair strand))
Description
FAQ
Comments (10)
As a Gagologist, I tend not to have interests in blindfolding...except in this exact case.
I've seen this just one time, a very long time ago. I was never able to figure out the exact 'search terms' to find it again. Didn't know there was a name for this (But also, I may have seen it before the term was coined). The one I saw used thin electrical tape, so that the gag was more like a cleave gag than a fully open-mouth one.
I'll see if I can do anything with this. I owe you that much for reviving a distant memory and providing me the terminology.
issue with this one is it's difficult to not have the hair stick out as if not compressed by nylon, it is the lack of training data, available training data in low resolution, and no artistic depictions since its so niche. attempting to synthesize more training data in different mediums, which should hopefully make the next attempt better at being applied to different styles. burned a bunch of buzz training multiple times in order to get incrementally more diverse datasets, and finally ran out. publishing to recoup some of the cost and will probably try again once I can generate a more diverse dataset / have more buzz.
@FloorInspector I'll see if I can achieve anything with a different style, but for the most part it's pretty dead-set on realism. I've also tried getting it to generate different kinds of ties, but often find that it overrides whatever prompts and additional LoRAs I use with a simple, arms-down-and-barely-behind-the-back tie. So far, my solution's been to generate with one LoRA first for the bound part, then inpaint with over the head for the mask. The problem with that approach is that the head tends to shrink a lot while the rest of the body stays the same. I'm having more success doing the mask LoRA first, and then inpainting the body to reposition the arms. I'll keep plugging away at it when I get some more time.
@gagologist cool, any help is appreciated, this one is a tough one but maybe eventually it will get there.
Try setting your learn rate, both unet and text encoder, to 0.0006 and train using AdaFactor.
thanks for the tip, you got it bud. thanks for the advice!
@FloorInspector No worries. Make sure to test all epochs, sometimes the best one isn't the final one.
@AcanthAI thank you for the advice and the buzz, working on giving this one another shot, creating and looking for more training data as well. I'll tag you once I have something better uploaded. thank you again for the buzz, going to put it towards trying to make this better, of course.
@FloorInspector You put up v2 for 1.5 lol
@AcanthAI oops! thanks