2nd version of SDXL-Lora has been released!
I trained it with 64.800 steps. And i trained it with about 10k-ish images.
Better facial structures
Somewhat good cum view but i've planned to make it better. ( fingers crossed)
One of the best but NOT perfect penis training you may find in this LORA. You will get barely bad output on penis anatomy which is i am pointing about "fusing tongue with penis tip". That was really bothered me for almost 1 year. I think i've managed to beat it most of the time but i need better SDXL checkpoint which is i need to train it my own but i need support to buy Nvidia A6000 because i have NOT enough VRAM. Also i can finally have Speed-up for TRAININGS. ( why don't i use cloud training ? - It is because i don't trust cloud-training for my "Training DATA-SET images" and most of cloud training servers are NOT allowing to train NSFW content or restrictive about it.
I've removed few thing from earlier version but some of them still exist.
I've spent almost 1 week for making proper Dataset, editing tags and 2 days for local training and testing. If you can't or don't want to support but using this LORA, please at least send your outputs, follow my profile and comment what you like or what you didn't.
This is a NSFW model that is intend to produce Cum (sperm) on face, mostly mouth, front etc..
( it is experimental, early phase and there is no 00, 01, 02, published versions, those were NOT good, so i didn't release)
If you see, text on the bottom of your output, it is because of the BASECHEKPOINT that is NOT MINE. My Lora is producing BARELY text, website address.
I used around 10K images. I'm trying new approach to Lora training with experimenting of tagging, words, images etc.
Unfortunately, main stream SDXL BASECHECKPOINTS are usually BAD producing at NSFW images.
I recommend you 2 checkpoints that are better for NSFW scenes;
1- https://civarchive.com/models/133808?modelVersionId=324958
2- https://civarchive.com/models/167764?modelVersionId=253250
Those are very good AT NSFW scenes.
To use this LORA of mine, You shouldn't USE new DPM++ 3M or new DPM++ 2M SAMPLERS.
I recommend you to USE DPM++ 2M SDE Karras.
Description
I trained it with 64.800 steps. And i trained it with about 10k-ish images.
Better facial structures
Somewhat good cum view but i've planned to make it better. ( fingers crossed)
One of the best but NOT perfect penis training you may find in this LORA. You will get barely bad output on penis anatomy which is i am pointing about "fusing tongue with penis tip". That was really bothered me for almost 1 year. I think i've managed to beat it most of the time but i need better SDXL checkpoint which is i need to train it my own but i need support to buy Nvidia A6000 because i have NOT enough VRAM. Also i can finally have Speed-up for TRAININGS. ( why don't i use cloud training ? - It is because i don't trust cloudtraining for my "Training DATA-SET images" and most of cloud training servers are NOT allowing to train NSFW content or restrictive about it.
I've removed few thing from earlier version but some of them still exist.
FAQ
Comments (8)
Let me start off by saying this is not a complaint about your LORA. Because it looks great gonna try it out. But in general the cum looks kinda fake like white cream, which is not your fault but SDXL it doesn't seem capable of producing it realistically. I have the same problem with my lora, and all the other cum LORA's out here seem to have the same problem, Hope that this will change in the future,
Hey, thank you for comment. :) Yes, it is a challenge. I am happy about penis anatomy and better finger management about holding penis stage. About fake cum problem, right now i have solid solution about it but literally to accomplish it, it will take about my 1-month-ish to complete, polish and release. if or when i grab A6000 GPU, i can make it in half time with x2 quality because of VRAM restriction. But i feel your pain, it is SDXL base structure problem.
Some of them do better with 'viscous white liquid' or something to that effect.
I have yet to find a model that handles 'eye/eyes' as a location. Most of them just ignore it altogether. Not a fault on your part. Just wish it existed.
You mean that in prompt "cum on eye/s" ? if yes, it can be solvable with this amazing idea, https://civitai.com/models/279972/experimental-guided-training-of-color-associated-anatomy-of-the-pussy
@Reelai Seems like that could work, if a bit time-consuming. (Maybe if there still isn't one by the time I get around to learning how to train them.)
It's been around for more than 5 months already:
https://civitai.com/models/139621/cum-in-eye-lora-sdxl-deepmind
@hansolocambo Missed it since it's not in the online generator. Thanks.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.