This is my first attempt to create a Lora.
This Lora supports the creation of BDSM concepts with Device Bondage. This means people (mostly women, therefore this Lora is only trained with female persons) who are kept in complete immobility and open presentation of the genitals by various devices.
It is a somewhat specialised form of BDSM, which is why, as far as I know, there is no extension for it yet.
114 pictures were used for the training.
Trigger word: device_bondage24
If this lora meets with interest, I will expand and improve it in a timely manner.
Description
FAQ
Comments (2)
Its not badly done, but that cover image is doing you no favors. Spend five minutes inpainting her mouth back to normal. :)
This model works way, way better than most in this area, great job! You must have done really well curating the dataset.
Since you said this is your first LoRA I wanted to mention that I looked at the metadata for the file, and I think you might be confused about either trigger words or captioning based on the generated images you shared. I could be wrong if you trained it in a way I'm not familiar with.
According to the metadata the only tag in the training set is the word "device". Unlike embeddings the name of the lora or the lora file ("Device_bondage24") is only used to include the LoRA weights in the model set (<lora:NAME>) and won't be interpreted in the actual prompt text itself.
Since you have some other trigger words listed it's possible you intended to use captions, but according to the file they weren't included in the training. Any words you are prompting other than "device" are coming from SD's pre-existing understanding.
It would make sense that "Device_bondage24" would still "work" because the model was trained on "device" and it does have a core pre-existing concept of "bondage". The CLIP encoder sees "device bondage" and "device_bondage54" as very similar ("_", "5", and"4" get new tokens but it sees the rest of the words the same).