This is a very basic version. I just ran some stability tests, and the results exceeded expectations—or rather, the ZIMAGE model is incredibly outstanding. Even the base model already covers most of the required features.

Here is the ZIT model and another thing u will need, if u dont have ZIT model.
https://huggingface.co/Tongyi-MAI/Z-Image-Turbo

Back to lora.

This version still has many flaws at present. Due to my lack of understanding of some untrained components of the text encoder and the base model, and it seems that its text encoder is completely different from that of SDXL, a few issues have arisen. For example, the generated images tend to be ????, and the rendering of ??????? is still not refined enough.

This LoRA is more of a ?????? version. I will further enhance the training and deepen my familiarity with the model in the follow-up. I believe a much more refined LoRA will be released soon. Of course, it will be consistent with the aesthetic style of q7-remix-llustrious-CartoonStyle. Please stay tuned.

P.S. My RTX 5080 is struggling a bit with such intensive training, so the process will be a little slower.

#1220

I think I have some idea of ​​what's going to wrong. It seems the base model lacks an understanding of ???? content concepts, which leads to many problems when training it with ???? content. For example, it doesn't know what ??????? are, or what a ?????? is, which is quite bad. However, it performs exceptionally well in style replication. Next, I think I should try training it with SFW content? Or perhaps try fine-tuning the model?