Original link: https://sololo.xyz/model?id=588

This LoRA is trained on Z Image Base, but it can also be used with ZIT. All sample images were generated using ZIT.

In terms of facial aesthetics, I personally prefer the Zv1. In terms of facial consistency and stability, Zv2 doesn’t feel much different from Zv1, but both are clearly better than F2K9v1.

Based on my limited experience so far, training a LoRA on Z Image Base for use with ZIT doesn’t seem like a good idea. Using Z Image Base for text-to-image also doesn’t seem like a good idea. In terms of both generation speed and image quality, ZIB doesn’t really have any advantages. ZIB’s real value may lie more in the derivative models built on top of it in the future—much like how SD 1.5 served as a foundation back in the day.

For upcoming ZIT LoRAs, I’ll likely go back to training directly on ZIT. The training methods and strategies still need some adjustment and optimization.

I’m hoping that better ZIB derived models will eventually come along and replace ZIT.