TwinkCockXL (alpha01)
This is my second major publicly posted LoRA. This was a way for me to learn how to use a large language model (LLM) to tag a dataset to generate an effective LoRA.
This is a concept LoRA that produces a penis in SDXL checkpoint models that are resistant to generating a penis, or generally do not generate a penis at all. There was also a secondary focus to generate a penis interacting with clothing.
I consider this an alpha because its ... okay.... it works most of the time, but it is not perfect.
Unlike my other LoRA, this LoRA was trained predominantly with short natural language phrases generated by querying the cogvlm-chat-hf LLM model with the following questions in the following order:
"Describe this man's age and body type in about 20 words."
"Describe this man's hair color and style in 6 to 7 words and start the sentence with "His hair is"
"Describe this man's penis size, length, whether it is erect, semi-erect, or flaccid in 6 to 7 words and start the sentence with "His penis is"
"Describe the clothing this man is wearing in 6 to 9 words, always start the sentence with "he is wearing"
The tags were then checked manually, and individually edited to fit within the 70 tag limit of SDXL.
The primary activation tag "twinkcockxl" was added to all images. A secondary tag was added to the following subconcepts:
"betweenpantsshirtxl" - a young man wearing a shirt and bottom wear (pants, shorts, underwear), with a penis visible
"cockthruflyxl" - a young man with a penis visible through his fly
"cockthrulegholexl" - a young man with a penis visible through the leg hole of his bottom wear.
"nopantsxl" - a young man with a penis visible while wearing a shirt but no pants.
"noshirtxl" - a young man with a penis visible while wearing bottom wear but no shirt.
"fullynudexl" - a young man with a penis visible while not wearing any clothing.
The use of secondary tags is not required, and in limited testing the use of secondary tags only does not reliably generate the desired configuration consistently. It appears that both the natural language prompt and the secondary tags should not contradict each other for most consistent generations.
An example prompt could be:
"a full body shot of a young man in his late teens or early twenties, he is college sophomore, he is 19 years old, with a skinny and athletic body, with long sweaty undercut red hair, he has a heart face shape, his penis is flaccid, he is wearing sexy sweaty shirt, but no pants, running on a football field with a determined look kicking a ball, late afternoon, golden hour, he has a dynamic pose, dynamic shot, rule of thirds,<lora:twinkcockxl01d_alpha:1> twinkcockxl"
The items in green are descriptions that are in the LoRA. The other descriptions would be pulled from the checkpoint you are using.
In terms of penis descriptions, penises were classified as "his penis is flaccid" "his penis is semi-erect" "his penis is erect". Sometimes penises were described as "large", "averaged sized", or "small".
When present, images were also tagged with "he is holding his erect penis", "he is holding his semi-erect penis". Images were also sometimes described as "he is touching his erect penis" or "he is touching his semi-erect penis". Images were tagged with either "holding" or "touching" but never both.
Other notes: Unplanned tags that can generate predictable outcomes included "he is wearing a necklace", "there is sunlight across his body". "Shot from underneith" was used, which was a misspelling.
While images with the fullynudexl tag were not wearing a shirt or bottom garment, examples were sometimes wearing a hat, baseball hat (described as "he is wearing a cap" or "he is wearing a backwards cap") or footwear (socks or shoes).
While the large majority of images used to train this LoRA was twinks, depending on the main checkpoint, it can be used to generate other body types and age ranges.
Faces: faces were not masked during training, and so the LoRA will impact face generation and interact with the main checkpoint. If a "clean" face is desired, the use of adetailer, face specific model is recommended. As well, having a separate prompt describing the face and not having the <lora activation tag> is recommended. I have also found that using a wildcard for nationalities or ethnicities can also change the face, as well as specific face shapes, if this information is included in the main checkpoint.
On testing, inconsistent penis generation or deformed penis generation may occur. I have found that using a different sampler or more sampling steps than the original checkpoint recommends may sometimes fix the problem. I generate X/Y plots to find these configurations. Generally I have found that Adetailer for the penis is no longer needed.
There were approximately 5000 images (including repeats and flips) used to generate this LoRA, and the most represented (~20% each) were betweenpantsshirtxl and fullynudexl. The other categories each made up about 10% of the overall set.
Special thanks to @markury , @zellian and spencer for testing an earlier version of this LoRA.
If you incorporate this Lora into your mixes, I would much appreciate a shoutout or credit. I do this for fun and education, please do not sell or use this LoRA commercially.