This model has been archived until further notice due to an active license agreement.
This is a project that I have been working on since October 2022 with multiple previous releases outside of civitai. Currently trained on roughly 5000 hand captioned images at 768px resolution using 1.5 as the base model.
The vast majority of the dataset is real photographs, not art.
Full list of trained keywords: link
The model works best at a lower CFG value, I usually start at 3.
The "Training data" extras zip file contains a lot of example images with prompt data embedded, and txt files with all the keywords the model has been trained on.
I have a discord related to this project for anyone that wants prompt help, discuss things, contribute, or access releases earlier. https://discord.gg/MzEaugjyFA
Curating images is incredibly time consuming, and training models at this scale isn't free. The plan is to continue releasing larger and better models for free, but I would greatly appreciate anyone willing to contribute, both with acquiring images and helping with the training costs, however small.
https://www.patreon.com/NotStrooge
Description
FAQ
Comments (39)
nice!!
Thank you so much for including the training data! I'm looking forward to trying it out. I wish more people would share what unique key words work for their models.
I don't understand a lot of the "words" you have in your example prompts. Are things like "eha" "wol" necessary?
They are necessary yes, simply using English sentences to describe something isn't reliable enough when the same words could be used for multiple concepts.
For example "eha" which describes neck corset, If you would simply prompt "neck corset" to get that concept, you'd end up with normal corsets as well with no individual control over it.
These keywords are tested to have little to no prior meaning in Stable Diffusion, which makes them suitable to help describe unique concepts.
@NotStrooge Wouldn't it be much more convenient to use just a concatenation like "neckcorset"?
@GlossyMannequin It is not that simple because everything is "tokenized" into smaller sub words for stable diffusion so the actual meaning of a concatenated word and space separated can be close or the same.
@NotStrooge What about replacing some letters with digits, e. g. n3ckcors3t?
@GlossyMannequin Might work better than plaintext but I don't see an advantage over the current tested method where the short few letter token is guaranteed to have no overlap.
The solution to this problem should come from a better underlying architecture for Stable Diffusion that actually understands what we are prompting. All of this are hopefully just temporary band-aids on top of early and flawed technology.
@NotStrooge The advantage is that such tokens are easily learnable en-masse, and the user doesn't have to look up the documentation each time
@NotStrooge This may not be a concept in the tokeniser, but I think it could be useful if custom namespaces could be defined, with a default namespace corresponding to the set of standard token ids, but with a way to define and activate used defined namespaces that start empty. The existing separation of words into tokens could still be used as that's efficient, but the tokens would be associated with new ids that are unique to the activated namespace. So "neck corset" would be 6906, 40408 by default (and indeed space makes no difference in this case), but if in a new namespace, the id's might be 500100 and 500101, or whatever would be outside of the default range.
@NotStrooge I ran into that trying to create a LoRA of my friend's dog, penny. I tried pen_dog and pendog, and most of the images came out with a giant pen randomly in the mix somewhere.
I get the most out of this model by merging it with another, more "well trained" model. It's an awesome model as a repository of raw data to do a merge on. Every version though, I get better and better native results. Amazing job!
I've been doing the same, been doing a 50/50 merge with Clarity v2 and seems to work pretty well doing that.
@ExcessivelySalty Ohh... I haven't tried Clarity yet. Looks like I got a project tonight!
@tspre726 Which models have you been merging with?
Merging is hit or miss for me. I got it working decently with RealDosMix, and so so with Clarity, and I got it working with OangeMix3 Hard Core some what, but with a 50/50 blend you will lose stuff from both sides which is rather unfortunate.
I have experimented with differential merging but haven't found any models that are happy about it yet because weights merge in a bad way on a lot of models. I.e Dreamshaper has the need for a VAE or baked VAE, and using that VAE when merging in Bondage causes every person to look like a drag queen and not having the VAE is even worse. Would be ideal to merge with dream shaper, but no dice yet.
50/50 seems to be the best compromise, but some things in Bondage won't work because they got dropped out, like a lot of the rwrists rthighs etc tags may or may not work where the heavier weighted stuff tends to work better.
If you look at the training data, you can see a lot of the good stuff is also what has the fewest weights so when you merge you are likely to lose those doing 50/50.
In short, I'm new to SD, so still trying to understand stuff and python isn't my main programming language. So if anyone has any guidance on getting a good merge (specifically for realistic anime/3d stuff, I'd love to know)
@xabrol Ahh I'm pretty new to merging myself, but that makes sense why when I tried to merge with Clarity at 50/50, I'd lose a lot of the lower weights. For instance I tried reverseprayer, even (reverseprayer:1.8) and it still wouldn't show up.
This is the best 3D BDSM model I've ever seen. Collecting and organizing the training set is very tedious and hard work, and I admire your dedication to model quality.
BEAAAATTTRRIIIICCCEEE
Any chance to have this as a LoRa instead of a full model?
Unlikely at this point in time, this is a relatively large projects with a lot of different concepts. I haven't seen any indication that LoRA is able to handle that.
There is now a LyCORIS available
Mate, this model is so much better than the pre-existing photorealistic bondage models, awesome work. I especially appreciate how many tags you added and how clearly you documented them
Thank you, I take this project seriously and I am glad it shows.
Usage Notes / Guide
Just few notes since I started experimenting with this model, and it was giving me some trouble. So I thought I'll share.
Optimal Values
CFG is 3 or 3.5 (between 2.5 and 4).
Sampler: DPM++ 2M Karrass (or SDE or 2S Ancestral - 2Sa gives better results but is much more volatile, changing image strongly with amount of steps).
Steps: 25 seems to be sweet spot, but you can get good results with 15, and 20. For 2Sa number of steps will strongly affect composition though.
HiRezFix: Do not use unless you are scaling from 512x768 or 512x1024 UP! Overall best width seems to be 512 with small variations, with 2/3 or 1/2 ratio, but you're free to deviate.
Use the keywords!
Author very graciously provided a list of unique tokens/keywords. Download them and use as reference! uwe ftw!
Model Is Volatile
This model will not follow prompt really well, and seemingly minor changes to the prompt will cause big changes in image composition. It's not necessarily a bad thing, but as a result tweaking the prompt - even if it looked great in showcase - is a bit pointless. Even minor changes to negative prompt have great effect on result.
Your best bet is to craft one prompt with things you're looking for and do XYZ plot with CFG between 2.5 and 4 (every 0.5), steps between 15 and 25 (every 5) and a whole bunch of seeds to see if you luck out with a good picture.
Alternatively just set CFG to 3.5, Steps to 20, and let it continuously generate until you find something good.
Merging is no good
I suspect that because model is so sensitive merging it with other more stable models does not produce good results.
Most other models seem to like CFG scale of ~7 but this model produces awfully overexposed/overbaked results around 7, so it's hard to find good CFG when mixing it.
Maybe I'm doing something wrong though, if you have any success let me know!
Mind ENSD
If you're trying to replicate images from showcase note that a lot of them have ENSD set to 31337. If you're not getting expected results check if it's set.
That's all I know!
If by any chance you know of better way to get desired results, or something I wrote is wrong - let me know!
Regarding merging I've had pretty good results with the Merge Block Weighted extension, and another solution is to use a normal weighted merge that on the surface might not work well for prompting. But instead use it for style transfer using img2img+controlnet on the raw output from bondage-v11.
This is something I've showcased on my Discord as I can't exactly upload such experiments as example images for this model.
how does the keywords work?
@TheHG All stable diffusion models are trained on image+text caption pairs.
Keywords/captions are the words that describe what an image looks like when training. And thus they are the words you should use when prompting if you are trying to generate any of the specific concepts this model was trained on.
@NotStrooge Speaking of keywords - once again thank you very much, they are great help. Although would be useful to describe shortcuts like hbab better - especially how they interact with front/back of position.
For example I discovered that "ponytail" will always put the tail behind the model even if she's facing away from camera. So is pretty useless for generating a tail behind back (from perspective of a rider).
Also was wondering about appt - it has "ponygirl" comment but I was wondering if there's any difference to actual "ponygirl" tag?
Otherwise - once again - great work!
@spiochkrakow443 The keywords with a bit longer explanations are from images a community member helped gather and caption, I didn't have time to test if/how they work. But I made sure to at least include his description of them for people who want to experiment.
A fair amount of the images tagged with ponytail are from a behind perspective, make sure you also include keywords like "back" and "behind", and possibly add "front" in the negative to steer the generation into the correct direction.
ponygirl and appt, I guess that's just how it turned out, I decided to caption the ponygirl images with both. I usually pick a unique single token keyword for new concepts that I think might benefit from it. A keyword like "ponygirl" tokenizes into multiple smaller tokens that end up sharing meaning with other things, like ponies, and girls, so it can be useful to have a neutral keyword that only refers to the concept.
Feel free to join the discord if you want help with more specific prompts.
@NotStrooge Thanks for the explanation once again.
Based on that I'll experiment with ponytails yet again :D
It's good to know appt is same as ponygirl as you are right - a decent number (~10%) of images would just render a pony, especially in landscape ratios like 4:3 or 16:9
I'll join the discord, but I prefer to ask questions / post guides here as that way everyone can benefit :)
I'm a little new to this, how would I be able to download bondage_v11?
Thank you for your work
Hi,
The link to keyword don't work anymore. Could you reupload it, please?
Fixed
@NotStrooge Many thanks!
Anything similar out there?
If you read the description of the model and click the discord link I'm sure you'll find something.
@NotStrooge Thanks. Alternatively can I get access if I join your Pateron?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.