The LoRA is trained using the BF16 version of Z-image base. The dataset consist of 60 Midjourney Niji images created by my friend Stan Katayama (creator of SDXL Niji, one of my favorite models): https://civarchive.com/models/120765/sdxlnijispecial-edition
Z-image base is already quite capable at producing anime images, this LoRA simply aligns the model closer to the "Niji look", specifically Stan's take on the Niji style. It can be combined nicely with other styles that are already "built-in" into the base, such as "Story book", "Studio Ghibli", "Vintage anime", "Pop-art", etc. to make anime images with a variety of styles.
Description
This is epoch 9 (epoch 10 is slightly overtrained).
Stan Katayama Niji Z5 D32A16P1 2026-01-27 23:11:54
Trigger: stankatayamaniji5z illustration
Repeat: 20 Epoch: 10
LR: 1 Scheduler: constant Optimizer: Prodigy
Dim 32, Alpha 16
(* 20 10 60)=12000, cost 1080
Noise offset: 0.03 Multires noise discount: 0.1 Multires noise iterations: 10
stankatayamaniji5z illustration. Woman with long, dark, wavy hair, illuminated by reddish streaks, and piercing glowing red eyes is caught mid-air, angled downwards as she falls backward. She holds a black pistol in her extended right hand, aiming to the left with a focused, determined expression as her head turns towards the viewer. She wears a dark, tactical jacket over a dark top and dark shorts, revealing a thigh holster. Red energy lines and sparks emanate around her against a brightly lit, blurred interior hallway, illuminated by a strong, warm backlight. Sampler: euler
Epoch Loss
Epoch Loss
1 0.336
2 0.322
3 0.325<
4 0.321
5 0.313
6 0.311
7 0.308
8 0.306
9 0.306
10 0.304