Yeah, needs more training material for V2 :-)

@TesterFranz great job though!

Use this in your prompt at the top: (Her tongue is coated in white translucent milk with chunks of white translucent jelly). This lora also works great for img2img. You can paint white streaks were you want the jizz to be with your favorite photo editor and run it through zimage. Bada Bing! It also makes the faces more realistic if you have older images.

Actually, yes. Some of the generations have longish, smudged nipples. Might be the nature of the turbo model?

could also be you need more steps for inference as when you train it your undistilling it so you have to use more steps like a base model. needs more testing.

yeah.. just use "Conditioning Zero Out" from Positive CLIP Text Encode output to a negative input in K-Sampler, so you don't have to load the negative CLIP.

Refractory period.

Your steps must be high...z-image work in range 4 - 12 steps. 9 is best and standard...

lora puts you over the memory limit so you offload to system ram which takes 20x longer

@Gramelto3 Thanks, that's true, I switched to GGUF and it was solved

If you think this is exciting, just wait until the FULL Z-Image model drops. The Turbo is the light/distilled model. LOL

@EryxAtys Great times ahead bud!!

@simonishere wait for Z EDIT

LONG LIFE OPEN SOURCE!!!!!