Please if you generate something cool - post it. I’m checking this page like a kid waiting to see what people make of it :)
This is a V2 of my Hung Twinks lora which has improved quality and can actually generate better background.
Dataset was cleaned up and training took 48 hours on a 6000pro at higher res.
I personally think it is a huge improvement, if you like this lora, please leave a like :)
I inference on de-distilled model
Model does best with prompts written by QWEN VL, this is also what I used to caption the dataset.
Enjoy and text me questions if you have any questions
res-2m
beta57
cfg 3
30 steps
1440x1440
Description
FAQ
Comments (23)
What is a trigger word?😋
There is no trigger word, just prompt it in detail
can you post more info about how you captioned, if you used comfy can you please include your workflow
Hey check my profile it has v1 workflow. Change Lora and it’s good to go
For captioning I used Qwen vl 8B model in LM studio
What weight
I inference on de turbo with 0.85-0.9
If you need to use it with a character Lora there is a better way to do it. Let me know
couldn't get it to work well with my zit character lora unfortunately, any advice much appreciated
what i did is I merged my lora with the model and used it to train a new character lora
kmax002/Z_image_nsfw (hugging face)
Go to Ai toolkit, select same settings for the de-turbo. change adress to the one I wrote. It will download the model and use it to train your character.
You can then use my model to inference (the bf16, not diffusers)
@Aelyadriel Hey I guess it’s not relevant now with base out, but I have it on HF if you want. Also issue is you select de turbo and change address to mine
Kmax002/Z_Image_nsfw
For some reason, it seems that in the generated image, the penis was blocked and incorrectly damaged.
Can you elaborate?
I've tried numerous times but can't generate a normal penis shape—most results are incorrect. Even when I copied the prompt from your towel-wrapped image, the generated output still lacks a penis.
@BOOMMM Try inference on de turbo and check if lora loaded correctly. (In terminal)
@kmax002624 Thank you for your reply, I have found the reason, it is due to the English language.
It works perfectly if you train your own character lora with the OP model, make sure to use the de-distillated option for your z-turbo on AI Toolkit and set his model as reference.
Thanks for confirming
Did you bake in a watermark, logo? While the LoRA works fine, it always puts a watermark, copyright, web url in the upper left.
A lot of images had watermarks but they were all properly captioned so normally it should not randomly reproduce a watermark (at least it does not for me)
Indeed. I also get watermarks very often from well-known porn studios.
Does not work with SwarmUI.
Error message:
"No backends match the settings of the request given! Backends refused for the following reason(s):
- Request requires LoRA 'Z-Image/Hung_Twinks v3' but the backend does not have that LoRA"
Other LoRA files in the same path are working.
the file is actually a Lokr not a lora, I wonder if that can be an issue for Swarm UI
there's no way you are using 30 steps with a turbo model. you get color blotches. what's going on?
Won't work with ZIT, ERROR: contains unexpected module keys, it is likely that this Lora is for a different model : lycoris_layers_0_adaLN_modulation_0.lokr_w1