This lora has been trained as a test case for z-image turbo.
Share your generated photo would be helpful for me to improve this model :)
Description
used real images for training
This version is still under test...
FAQ
Comments (17)
Was this trained with other AI images?
How much does it affect other facial loras?
1.0版本影响不大，1.1很奇怪，影响有点厉害，不建议用
Keep getting asian faces with this, even when prompting for other races.
Maybe u need to try a little bit lower weight. I've tried to use different races in v1.1 dataset but failed to achieve an acceptable lora. That may be associated with Z-Image Turbo has been distilled. I'll try for different kind of datasets, keep following for upcoming updates
What about providing some images for the dataset?
Seems to be too heavility trained that overhauls defaut images' style.
v.1.0 works best with 0.4 weight.
Thanks for your feedback, I think the datasets may need to be improved
I would try to train a slider. It may helps...
Good job, but i like v1 better, v1.1 is altering face too much it seems.
This works fairly well with a small negative strength to give the women of Z-Image a more natural look when they tend to be a bit exaggerated in that area by default.
Using this Lora v1.0 makes faces look Asian. The adverse effect on faces is especially pronounced in images with two or more people. For example, without Lora, it can output Asian and European people side by side, but using Lora causes it to ignore the prompt and make both Asian.
To be honest, the breasts are drawn beautifully, but the face lacks the natural Asian look that Z-Image is known for and instead feels more like an AI-generated result. Still, I’m looking forward to future improvements. I’m rooting for you.
it might be a good one if it wasn´t trained on faces