style_by_zeekzag

We will delete it upon request.

This is a style LORA for furry artist zeekzag.

While creating this LORA, I discovered that

50 to 100 training images are sufficient for creating a style LORA;

in fact, having hundreds of images seems to reduce reproducibility. In fact, comparing a LORA trained on 200 of zeekzag's illustrations to one trained on 100,

the latter achieves higher style fidelity (needless to say, this LORA was created using 100 images).

Depending on the style you wish to replicate, the variety of styles, and learning rate settings, 100 images might sometimes be insufficient. However, it appears that more training images do not necessarily result in a better LORA.

Translated with DeepL.com (free version)