V3 UPDATE:
This update greatly expands the promptable camera angles and adds or improves several other concepts. Here are some triggers to try:
Poses:
squatting,kneeling,all fours,lying on stomach(no lying on back yet)Camera angles:
from front,from above,side view(stricter),from side(gives a wider range of side angles)Penises of Color:
brown cock,dark brown cock(you sometimes need to add e.g.dark-skinned maleorblack manto avoid a default-tone body, and titrate strength until it makes him the desired shade without effecting his blower)sideways blowjob, cock in cheek, cheek bulgefor chipmunking/cheekjob.chipmunkingwill just give you chipmunks.A bit of 2boys or 2girls training, recommend
kissing a cockorlicking a cockinstead ofsucking a cockfor multiple women, no consistent solution yet for e.g. one on shaft and one on balls.
Future concepts to add include upside-down, on back, kneeling over, feet in the air, gloves, from behind, and more. Show me what you got, and let me know what additions you're thirsty for! Much of this update was driven by feedback and requests.
---
Want to get POV BJs from any model, in any style? Need adjustable strength to work alongside your other LoRAs and embeddings without deep-frying your gens? Well step right up and put a cock in that mouth, because the POVBJer does just that.
Take any character you've trained and add this LoRA to get them going down. Trained on all professional adult actresses but generalizes very effectively to men as well.
Works great on most models and both converts anime to realistic and realistic to anime. Of course, works great with no character added as well. Effective at nice low weights of 0.2-0.4.
"Redo" because this was re-trained from scratch to eliminate non-professional content, since I found that my personal LoRA was working better for me than the models and embeddings I found on here and I wanted to share the love.
Description
Huge expansion of dataset to 153 images to add or improve several concepts (including hands/fingers) with a goal to maintain strength and performance on POV images. 153 images X 25 repeats X 10 epochs. Trained with 0.05 denoising, random horizontal flip, .00005 Unet and .000025 TE LR with Lion optimizer.
FAQ
Comments (14)
No matter what checkpoint I use this with on Stable Diffusion this lora produces weirdly disfigured details throughout the image. I'm not sure what I'm doing wrong, as it looks like others are getting great results.
It seems quite heavy handed when using with other loras but reducing the lora weight helped for me try 0.65
Dunno if it's the models I'm pairing with this, or your model, but I always have the bottom lip over the dick.
CLIP thinks that dicks are tongues, so if you put "tongue, tongue out" in the negative prompt that helps push the model's attention away from the base model and toward the lora's data to keep the anatomy correct. It also helps if it makes glanses that look like tongues. It should only be needed when running non-NSFW checkpoints or running the BJ lora at much lower weights than other loras in the prompt. Note, there is some training data in there for tongue out, cock on tongue in the side views.
瑟瑟素材/成果/技術交流分享
NSFW test and share
https://discord.gg/4V9FUdjqqU
Doesn't seem to look great with any character Lora, it just give it the style of the blowjob lora ruining them. Works great with no character lora though
just set minimum lora weight for any character. e.g 0,2
tq
@Ryuzhal I've been able to get it to work with basically any character lora. Can you share some example loras you were trying to use it with, and I'll see if I can make good examples with them?
@honk_honk I wonder if I accidently used a different LoRA when I tested this. Coming back to test it seems to work a little, I'll get like 1 in 16 gens that looks decent enough and even then gotta inpaint a lot.
Z lot of the time it won't generate the man. I couldn't get POV on stomach to work at all. This my my prompting
masterpiece, high quality, highres, (1girl, 1boy:1) nsfw,
<lora:Zelda-v2-01:0.7>, BotwZelda, blonde hair, braids, long hair, hair ornament, one eye closed, one eye open, (blush:1.5),
indoors, bedroom, (looking at another:1),
blue bra, blue panties, blue lingerie, large breasts, mature female, sweat, <lora:sweat_lotion_v2:0.2>,
man to side, (squatting and sucking a cock with a cock in her mouth), blowjob, from side, arched back, <lora:POVBlowjob_v3:0.3>,
If I turn your lora down I get shit gens, if I turn it up it's looks terrible with char lora, dunno
Please retrain this LoRA and add proper keywords to each dataset. They are already separated but there's no proper keyword to call any of them so my results was disfigured mess.
Sure you'd get more consistent results with separate models for each combination of pose, camera angle, outfit, ethnicities, etc. but you'd need many more good source images to accomplish that. The images are roughly categorized but there's considerable overlap between the categories. I'd consider training some sub-models if there's a good tool to pull only the images with specific tags automatically.
You still shouldn't be getting a disfigured mess often, though. To troubleshoot, start with a low lora weight like 0.4, use a good negative embedding like bad-picture-chill, start with a nsfw-capable realism model like Cyberrealistic, and get everything working at 512x512 before moving up.
@honk_honk thank you for the reply and tips. sorry if i sounded too rude before.
I’m making a request for a LoRA, where the girl is taking the penis out of the open zipper of a guy wearing pants. Can that be done?
love how when I search for blowjob, or fellatio this never comes up and I see it every couple of months on civit and use it one or twice. oh yeah I got that lora. like bro most of us have millions of lora files. remembering this particular one isn't gonna happen. just tag it after the name *blowjob
Details
Files
povbj_scopecreep_b1-000010.safetensors
Mirrors
povbj_scopecreep_b1-000010.safetensors
redoOfPovbjer_v3.safetensors
povbj_scopecreep_b1-000010.safetensors
povbj_scopecreep_b1-000010.safetensors
redoOfPovbjer_v3.safetensors
povbj_scopecreep_b1-000010.safetensors
redoOfPovbjer_v3.safetensors
povbj_scopecreep_b1-000010.safetensors
redoOfPovbjer_v3.safetensors
redoOfPovbjer_v3_BLOWJOB.safetensors
povbj_scopecreep_b1-000010.safetensors
bojvop.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.