Illustrious knows all about bondage, bless it. But it does seem to struggle with the idea of someone tied/cuffed to a radiator. So this is an attempt to address that.
Description
Doubled the training data, and added a couple of epochs in the hope this will strengthen the effect.
Trigger remains "radiator_bondage", but as i think some of the difficulty it was having was because rope/cuffs were mixed together in training images, have added extra tags for
Tied to radiator
Cuffed to radiator
Chained to radiator
to help differentiate.
FAQ
Comments (7)
https://civitai.com/posts/14808224
Let me share my result and follow you for future contents!
Really glad to see this!
I'm not yet convinced of the results, but this has a lot of potential. I'll definitely be keeping an eye on this.
As for finding training data, I feel the pain. I think that is probably the primary reason why no BDSM fan has tried making a LoRA for this particular topic before.
Thank you - and yes, i agree, this isn't where i want it yet. Will start feeding it any good gens that come out of the current model to try and improve it. The trouble is that in crappy screen-captures the actual point of contact with the radiator isn't always that clear, so it thinks "a girl next to a radiator" is fine...
@SimianHose Thanks for the reply. I see.
I also tested this quickly, to get some intuitive idea of what it does well and what it doesn't yet.
The LoRA already often gets the pose right. This is very nice, as the poses relevant to this BDSM trope pretty much only appear in this one BDSM trope.
What I see as a major challenge is to get the background aligned right. The current version (v2.0) rarely generates any pipes or other suitable places for attaching the handcuffs anywhere near the character. Also, having both wrists cuffed to the radiator (using just one pair of handcuffs) is much more challenging for the model to draw than the case with just one wrist, because as we know, handcuff chains tend to be short, so the positioning needs to be perfect.
Adding in synthetic data is indeed a good approach. After reading Shumailov et al.(2023), I wondered for the longest time why synthetic data augmentation works at all, until Gerstgasser et al. (2024) published their (AFAIK much less-known) counterpoint study. It seems that using some real data is enough to ground the model.
From my first couple hundred renders with this (to get an idea of the captured distribution, pun intended), I don't yet have any images worth publishing, but I did get a couple of promising base images that will go into my inpainting queue. If I get around to it (from all the other stuff in the queue, lol), I'll be sure to add them to the gallery, to provide some synthetic training data.
This is probably my second most favorite BDSM trope, after chairties. I wonder if I'd have any images lying around from the days before GenAI that could be added to the dataset. I think I only found a couple of them, ever. This trope if anything is pretty niche.
@SimianHose First two images detailed and posted, here and here.
Personally, I think the bondage looks better in the first one, but the pose is better in the second one (more relaxed, implying she has been sitting for a while).
I still have a couple more promising base images from this set, but on an AI timescale, these are currently slow to make (about one hour of inpainting each), so I might concentrate on other things first. It might also be good to work with different characters and outfits, for variety.
It would be nice to get more creators to join in.
@Mathemagic Thank you for sharing these, and for taking the time that was undoubtedly necessary to perfect them - very much appreciated, and great results. Completely understand that there are other things to do - will fold these into the next training run. Thanks again for your interest and support. Maybe I should try a bounty, see if that inspires some more potential training images... 🤔
@SimianHose Glad to hear you like the renders. As mentioned, I might make some more of these when I get around to it. I love this BDSM trope.
Personally, I don't care about bounties, but others might, so maybe worth a try.
And completely unrelatedly, I've been meaning to comment on your observation that everyone wants to make AI art, but no one wants to look at it. I have some hypotheses.