My first Z-Image test. Since it's distilled model the results aren't perfect, but acceptable...

Trained on carefully generated, edited, tagged and completely anonymized dataset.

My main focus train on such a way that it only affects the related parts of the image. It does great with ?????? sizing but sometimes fails with other details...

Here is example prompt to create similar images:

bimbo woman. A woman with long, blonde hair and prominent tattoos on her left arm and thighs is photographed in a black, glossy latex top and matching bottoms against a gray tiled wall. She wears a black choker with a silver heart-shaped ring and a silver cross pendant. Her top is a bustier style with a central silver ring at the bust, and her bottoms are a ????? with a silver buckle. Her body is large-framed with very ?????????????. She is standing with her left arm partially visible, her head tilted slightly to her right, and her gaze directed at the camera. Her lips are full and glossy, and her makeup includes dark eyeshadow. Her right arm is out of frame. The top is shiny and reflects light, contrasting with her skin tone.

I used official comfy workflow. Just used RES4LYF res_2s sampler with 4 steps, but these default should do ok too:

0.6-1.0 strength
Euler/simple
CFG 1
8 steps