V2.
ALSO USABLE OUTSIDE CARS (whit a bit of prompting) IN POV, 3/4
(penis in trained one direction it wont turn)
I retrained it focusing more on the penis for cleaner results. It’s trained in BF16 over ~5,000 steps on 55 images at 32 epochs. The LoRA is meant mainly for POV from the driver looking down inside the car. It can do outside-the-car shots, but that part isn’t fully stable yet.
It handles day and night fine, and works best in horizontal or square formats. 9:16 vertical tends to give more unrealistic body positions for now.
It behaves better with short, direct prompts. If you over-complicate the prompt, it starts mutating on the sides because the pose and environment are already tricky. It also plays decently with character LoRAs, but you have to gauge the weights.
All the examples I posted were done on a simple workflow: no low-res pass, no upscale, no refiner. Just straight renders at CFG 1.1, 8–12 steps.
Model-wise, it works on several bases. Euler Simple is the safest. Personally I prefer Res2 / Beta 57 for the crisper, more realistic skin tones, even if they’re a bit trickier to control.
I’ve tried to make this LoRA as versatile as possible, considering it’s a complicated pose in a very tight, closed environment. That naturally limits flexibility, but it still holds up surprisingly well. I’ve tested it in scene of the 50's , 60's and even though all the picture of the dataset were from this era it help surprisingly well, and it generally follows along without falling apart, so it’s not too bad.
I will retrain in the future but I'll move to other Lora for now .
Trigger: BRKN_CARBJ, car blowjob,
Prompt exemple:
BRKN_CARBJ, car blowjob, POV of The man sitting in the driver's seat looking down toward his crotch where the woman as her head on his lap. His large penis is coming out of his pants, and she as in her mouth, giving him a blowjob. Scene takes place in a car interior, sunlit day, playful grin, intimate moment. We only see the man' s bottom half.
Cute blonde 20 years old, small breasts, slim and fit, naked.
Description
This LoRA was trained on 50 images over 5,000 steps and is still very much an experimental toy. It’s built to be flexible: it holds up nicely in both daytime and nighttime scenes, and in my tests it doesn’t overpower or distort the base character. It’s especially fun in close, intimate setups like the front seat, where you can really let the mood and lighting do the work.
Play with it, break it, push it in different directions and tell me what you think. If it behaves the way I hope it does, I’ll happily cook up a few more in the same spirit.
FAQ
Comments (18)
dick are still ugly, but thanks for your job :)
yeah I will redo it exactly for that reason , well mostly
yea i don't know why train nsfw in ZIT is being worse.. I also train it but look ugly too, train with 3000 steps and 200 dataset image, would you like to share how many batch and accum grad, also steps you used to do this please? Thank you <3
@nongreiprompt188 I think 3000 Steps might be to little for 200 images. 3000 was good for my initial Training for a NSFW prototype but i only used about 25 images. I might be wrong but lets see. Currently i have one running with about 60 images and am a 5500 steps. The results are still getting better with the iterations and i think about letting it run just to see what happens
@Anatomical Wow thanks for info, maybe i need to increase the Steps.. are you using like high batch? like 4/8/12?
@nongreiprompt188 I dont have a great GPU so i cant use batched training. Usually i just let it run through the night, check results on the day and make changes to continue at night.
@Anatomical I have a bad computer too haha so i'm using online GPU services, it cost me a lot yesterday when i'm using 4090.. Idk if 5090 can make the training faster and with lower cost. however this is interesting.. Let's work together to solve this mysteri hahaha
@nongreiprompt188 haha, i use my own 4070s and get about 6sec/it. 5090 should get you about 1.5-2 sec/it. So it would be more then 3 times faster as my training right now. I personally am just to lazy to setup a runpod or some other online GPU service
Add the penis helper lora and you'll be fine
Some of the dicks in your examples look quite good. Its really good tbh. The balls and tip are the hardest part and if you have recommendations about getting/preparing good data for it, i would be really happy.
How did you manage to get the original character not changed that much? I am currently working on a dick lora but my prototypes change the original generation to much.
Maybe i need regularization but i have not done it with it as of now.
I'm playing with it atm and by tweaking it a bit I'm getting decent results
yea the 3rd examples look perfect, wondering what make the dick look weird? it is because of the base model hasn't released yet? or because our setting not yet correct with the model?
@NUVoize what do you think about the best GPU to Train Lora? i still can't consider beetwen RTX 5090 vs 6000 ADA
"and she as in her mouth" in the example prompt. that's a bit of a mess
wasn't by accident
Man, half these LoRA trainers can't even spend 30 seconds to type a prompt if a gun were pointed at their head and we're supposed to be able to expect them to curate data and write decent prompts for their training material lmao.
@saturnalis This is why I mostly stopped using random LoRAs a while back and started making my own. Early in a model’s life it’s always tricky, because every base model reacts differently to LoRA stacking, training detection, and all the tiny details that decide whether you get a great LoRA… or waste 5–10 hours trying to force something the model just can’t do at its core.
Z is a very small model, so from what I’ve seen it’s easy to bend but it falls apart fast if you push it too hard. Still, it’s a really promising base. I’ve got a bunch more LoRAs trained and testing right now, and some of them are honestly pretty interesting, this model keeps surprising me.
The way I work, I value flexibility over perfect exactness in the LoRAs I share here. One-trick ponies are easy to make, but they usually come with so much pollution that they’re only really usable for that one specific image. But let me know if you have trouble with this one I have about 50-60 prompt for it and my Z-Turbo prompter is pretty much ready.
can say it looks great from first tries, what do you recommend for the weights while using character LORAs ?