I’m still very new to training LoRAs and this is also my first proper dive into working with the Z-Image base models, so this has been a big learning process from start to finish. A lot of testing, retraining, caption tweaking, and comparing results between base and turbo, and I’ve already learned a huge amount from it.

The goal with this version was to improve consistency, reduce deformation on lighter prompts, and get better stability across different lighting and environments. I’ve included a few different scenario samples to show how it handles modern settings, dramatic lighting, and stylised prompts.

I’d really appreciate any feedback, especially if you test it with different samplers, or workflows. If something breaks, overcooks, undercooks, or behaves differently on your setup, please let me know. That kind of feedback is incredibly helpful and will directly shape future versions.

Huge thanks to everyone who takes the time to download and experiment with it. One of the things I love about this community is how much knowledge gets shared around, so I’m hoping this contributes back in some small way.

Looking forward to seeing what you all create with it.