Triggers with yoshida hirofumi
, weight 1 works best with most models.
Used Cetus-mix for example images.
It was really a journey to make this LoRA. For such an easy design, it was actually complicated to make it work. It took me 13 tries to get it close to perfection, but well, I think it was worth it.
General tags: yoshida_csm
works fine already, but you can use 1boy, black hair, black eyes, mole under mouth, short hair, crossed bangs
if necessary.
For Yoshida's gakuran, just use gakuran, buttons, black pants
, sometimes you need to put green pants
in the negative prompt (?).
For Yoshida's casual clothes, use official alternate costume, open jacket, white jacket, black pants.
More stability and quality than v1.0.
Changed the training parameters.
More flexibility, can easily change clothes.
Trained on the old version of kohya-LoRA-dreambooth on Google Colab.
Trained on a set of 35 images.
10 epochs, 2 batch.
8 repeats
0.0001 LR + 0.0001 unet_LR + 1e-05 text_encoder_LR.
cosine_with_restars
scheduler, 0 lr_warmup_steps.
64.0 net_alpha, 128 net_dim.
Training resolution 896x896.
Model: NAI Model (animefull-final-pruned.ckpt), no VAE.
225 max_token_length.
fp16 mixed_precision.