Initial.

For v0.0.0, I used 93 training images, primarily consisting of images from 2016 through 2025 on Danbooru.

Below is a comparison of the training configurations with the already-released Eisu (Eith) LoRA by StellaDeCorbin.

Eisu (Eith):
dim20/alpha20, AdamW8bit, UNet LR 1e-04, cosine_with_restart[*1] 4080 steps, warmup 100, 34 images (1024²)

v0.0.0 (this LoRA):
dim32/alpha32, CAME, UNet LR 3e-05, cosine 930 steps, 93 images (1024²)

[*1] The number of restarts and the details of the training images used by StellaDeCorbin are unknown.