fishine / hirame / ireading62 style LoRA (PonyXL)版本v12 (ID: 479341)

This model was trained using image from fishine /hirame /ireading62.

I made a few iterations until I got a stable model that had the intended look. The model used for training was pony v6, so models derived from it should also work.

Note: I like the style when using halftone, hatching \(texture\), jaggy lines

Changelog:

• v12: first model published (12th iteration I've trained)

• v14: removed some images which were causing some instability; increased gradient accumulation

• v15: relabeled images with wd1.4 tagger v3, using 0.35 threshold; seems a bit less stable, but more flexible

• v22: recreated the dataset (total 121 images); some images will need manual labeling since the model quality seems limited by the wd1.4 tagger (v2/v3)

描述:

Trained with 16 gradient accumulation steps, for ~192 epochs on 49 images (labeled by wd14 tagger)

训练词语: