This is a Lora that integrates various erotic concepts without watermarks(No need to add "watermark" or "signature" to negative prompts.), maintains flexibility, and adapts to commonly used resolutions.
03/20/2024
generation data has been added to the example image
It wasn't until someone left a comment saying they didn't know how to use it that I realized the GENERATION DATA was missing for those images.
Now you can view typical trigger words through it. Or you can just drag the images in using comfyUI.
03/13/2024
In order to get smaller files, all the concepts were put together for training.
Maybe you will feel that a concept works differently than the previous version.
Add:
full nelson
lesbian fisting（Hard to trigger accurately）
reverse cowgirl
ride titfuck
upturn face fuck
vagina speculum
Back Riding
Change:
"drinking pee" The trigger word is changed to “piss in mouth”
The concepts may not be described precisely enough, see the example pictures for details
Concepts included in previous versions are still available in the new version.
The .zip file is the tag form used for training, and can be downloaded if you need to know more about trigger words. *this table has the proportions of each resolution image in training, which is helpful for concepts that are difficult to trigger.
completed
Fingering Pussy(Other guy's hand&self)
piss in mouth
fucking machine
gag bondage (ring gag&bit gag&ball gag)
rope bondage
tiedbreast bondage
BUKKAKE(*Optional trigger word:multiple penis)
deepthroating dildo
deepthroating pov
blowjob side
lying anal
Sitting anal
doggystyle
standing(vaginal sex&Anus sex&none)
lesbian fisting(See zip file for trigger words)
full nelson(vaginal sex&Anus sex)
reverse cowgirl
ride titfuck
upturn face fuck
vagina speculum
Back Riding
Recommendation: if using Juggernaut XL, use a low CFG of 4, 4.5 or something like that.
Next
Slowly building a new dataset (expect someone to provide images that meet the criteria, or tell me what concepts are missing right now, as long as I can find enough of them, of course)
Testing. Balance the number of images available for each concept and try simple tags that do not conflict with each other for training.
If the testing goes well, or if the quality of the dataset images improves, or if there are enough images for new concepts to be added. I may try to train directly on someone else's realism checkpoints, and if the final quality is satisfactory, it will be released as checkpoints.
Face Restoration XL - v1.0 | Stable Diffusion Checkpoint | Civitai
(Completed)Face Enhancement Specialized Models. The odds are that it will use checkpoints that others have allowed to be merged and released. Add data on common face angles in various pornographic images in a lora fashion and differentiate between female ethnicity and age.
Thanks for their dataset support!
Description
add
lying anal
Sitting anal
doggystyle
standing
FAQ
Comments (12)
Does anyone know how to avoid shared vocabulary in different concept labels from influencing each other?
Trying to use natural language annotation, but those shared vocabulary in the concepts added later affects the concepts that were done before the training.
can you clarify which versions we should download if we want the full spread of capability without interference? Is it all the non-Full models or what?
Only the most digitally maxed out non-Full version is needed.
hey bro,
why a 5.5GB Lora ... omg ^^
Because the datasets are still in the process of being collected, they are all trained separately and need to be merged and extracted.
When the images are organized and a lora is trained it won't be such a big file.
i see...
my new model v3, ist very sensitiv on your lora, at least the new stuff you have implement
@sevenof9247 Now the file is smaller. I'm looking for a checkpoint suitable for training pornographic content. Do you have any suggestions? After testing for so long, I found that small batches of images are not suitable for using natural language tagging at all.
@suede2031691 its not really better to train porn on a porn-checkpoint, you sould first check if the model include some stuff you want to train or what images are generated allready from your prompt.
try my ;)
@sevenof9247 Okay. Now there's time to start trying. The goal is not porn-checkpoints, the focus is on nude structure and facial realism.
@suede2031691 try BetterThanHentai XXXL
@suede2031691 https://civitai.com/models/162103?modelVersionId=363977 a very good model, not sure if author wants training
The quality of the concepts added subsequently is getting worse and worse, so next I will reorganize the dataset and train it all together to get a lora with a smaller file.