Z-Image Turbo "de-distilled"

This is a LoRA for fun. Until z-image base model released.

"de-distill" a distilled model is a stupid idea, but ... still no z-image base.

¯\(°_o)/¯

Now I can use CFG (and negative prompt) again, like a normal model.

Better details? Higher contrast? Seems a little bit.

Technically there is no such thing called "de-distilled". What the model learned is learned. This LoRA just make the "distilled" model act more like a normal model so that using CFG won't burn the image. I don't know how to name it.

Trained with ~10k images with lr 0.00001. 1 epoch only.

Technically this is also a "adapter" and you can train LoRA on top of it.

Differences with ostris's adapter:

• real images, no repeat, rather than self-generated images

• less trained (10k steps vs 57k steps)

Haven't tried to training a lora on this yet. I just want CFG for now.