Z-Image Base finetune trained on a moderate dataset with very heavy focus on female form and photography. The dataset contains multiple photographer styles, but in practice "photograph" probably works as quality modifier, ie. "photograph of" is going to give better output. Watermark cleanup was not complete so unfortunately there will be some hallucinations of those.

The anatomy is still not quite there at 50 epochs in, but I would think that if you slap on some other anatomy lora on top of this model the results should be more stable than using the original base. The model likely does not do particularly well with male nudity, as it is of passing interest to me and I do not have dedicated dataset for it. Model mostly knows what it knows in relation to interactions with women.

The samples for fp8 are generated with Fun Distill lora for 4 steps @ 0.7 strength with 5 steps, but refer to individual samples for detail. The samples for bf16 are generate with Fun Distill lora for 8 steps @ 0.75 strength with 9 steps. I have only briefly tested generating non-distilled but the results are passable from the short test.

The v1.5 is probably way overtrained and is kinda annoying to use, but I thought I will share it anyway. Output without using Heun is very grainy, but your mileage may vary depending on your chosen workflow. It's better at adult concepts than v1.0 but still far from perfect. Samples are generated with 4 step distill lora @ 0.5 strength with 10 steps, using DPO lora with low strength (0.25). V1.55 seems generally superior to v1.5, I'm using 8 step distill at 0.75 and 10 steps for samples, with low strength DPO lora.

I'll add list of photographers later, but you can spot some samples in the generated images.

Distill loras: https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill

DPO lora: https://huggingface.co/F16/z-image-turbo-flow-dpo

...Low Optimization is just a name, nothing special.

If you want to help pay the electricity, you can do it through https://ko-fi.com/konflikti