I Trained a Conceptual Anime Style LoRA Using the Original Data from Z-Image Anime
Hi! I used the original training data from the author's Z-Image Anime / z_anime model to train a style LoRA, and the results turned out very well. Thank you for creating and sharing such a great model and dataset.
About the LoRA
This is a conceptual anime-style LoRA trained for approximately 6,000 iterations.
The LoRA mainly changes the overall visual style, including linework, coloring, shading, and character rendering. It has very little influence on the original composition, pose, camera angle, or scene described in the prompt.
You can safely increase the LoRA weight to 1.0. In my tests, it remains stable at full weight and does not show obvious signs of overfitting.
Comparison
Image 1: Without the LoRA
Image 2: With the LoRA at weight 1.0
The examples below show the LoRA applied to different characters, poses, compositions, and scenes.
LoRA Page
You can find the model, more example images, and usage details here:
https://civitai.red/models/2766487/qyzanimestyle-lora
Thanks again for your work on Z-Image Anime. I hope you find this LoRA and the comparison results interesting!