Like the original this is "non face changing", at least with the idea that some people are familiar. But, in fact, the face is quite different when using it vs not using it but the essence of the fictitious people used in the training set do not bleed through since none of their particular characteristics were maintained.

So, effectively, if you have a character LoRA enabled then this LoRA should not alter the face characteristics in any way that obliterates the likeness.

A sort of masking was used to eliminate these details and the captions reflect the data meant to be presented when prompting and, as a result, the image tends to lighten up when the LaRA kicks in since the masked area became part of the training (white areas). Thankfully, however, these areas don't seem to bleed through like some other LoRA I made so the effect has been maintained from one trainer to another, which is kind of a odd way to describe it since I went from ai-toolkit to OneTrainer O.o, the first version being trained with ai-toolkit.

All of the images in the training dataset were generated using AI, 56 images were used and 56 different prompts were used for each image, in order to make sure the fewest, unrelated features were repeated (preventing unrelated content from being learned).

It would appear that Z-Image wasn't trained on the word "" so using it, in my experience, allows for more character versatility, which I did for training in the captions.

With regard to the version differences, from v1 to v2, aside from having used a different trainer, the dataset was slightly tweaked and the captions were altered but the result remains basically the same, except that I find more versatility with this version and, despite the additional steps, it seems that this version is less prone to overreach allowing for a wider range of weight without huge impact.

The LoRA makes its appearance at around weight 0.2 but doesn't fully realize until 0.3. My experience has been that 0.35 can be a lowest operating point when mixed with other LoRA, allowing for Z-Image to keep its "real" characteristics. Beyond that, I'm not particularly interested in the "ultra-real" sort of thing so I tend to push my LoRA to a 3d effect, illustration or just anywhere a bit more colorful than what one might expect to see in any average photo so I roam at around 0.6 to 0.8.

Hitting 0.8 and beyond will introduce some unfortunate trained artifacts such as bad lower teeth since there were some data that couldn't be left out of training, that being one of them. Additionally you'll get that Illustrious feel since most of my training data is sourced from a refine of that effort, though some realistic LoRA were used for part of the set.