Watermarks, text, etc weren't pruned in this version. Some useful negative tags may be: "watermark, artist name, english text, korean text, patreon username, patreon logo"

• 530 images

• Tagged with wd14-swinv2-v2

• x768 base resolution with aspect ratio bucketing

• 20 epochs, batch size 1

• 1e-3 cosine LR (matching U-Net and text encoder)

• 256 dim

• Clip level 1

• Anime-focused base model

Some mistakes were made in the training of this model but the effects can still be positive depending on the clip level and weight given to the LoRA.