Not recommended for realistic models
Main tag : si \(arknights\),
Secondary tag : dress, hairband, gradient skin,
Training LoRA Powershell info :
train_speed_mode = fast
resolution = 512
Batch size = 2
network_dim = 16
network alpha = 16
mix/save_precision = BF16
optimizer_type = AdamW
optimizer_args_other = foreach=True
GPU RTX 4060Ti 16GB
Torch 2.9.0+cu130 + SDPA+ gradient_checkpointing
compile_mode = max-autotune-no-cudagraphs
Description
ตรวจพบ - วันที่ 23 / เดือน 7 [จะแก้ไขภายหลัง]
Unet ไม่เสถียรขณะใช้งาน LoRA ซ้อนๆกัน
รายละเอียดขนาดเล็กหายไป
FAQ
Comments (1)
Thank you for sharing the great quality LoRA!!
I have a question about training.
Your character LoRA can switch outfits, and when generating using prompts for other outfits, the default outfit’s shape, color, and other features rarely bleed into the result.
For this result, do you include images of other outfits in the dataset?
And, how do you go about editing captions? In particular, do you include both the character appearance tags and the costume tags? Also, do you include natural language descriptions?
Sorry for bothering you. But this is important for me. thx.