概况

基于animagineXL v3和kohaku delta训练的blue archive的plana人物lokr模型，也能用于出arona。

• 触发词plana\(blue archive\),采用详细标注的方式打标训练，因此想要完整还原人物特征时需要尽可能的添加相关特征的提示词；

• 由于训练集本身混有一定arona的图片，因此也能用arona (blue archive)来出arona，但是效果上没有那么优秀；

• animagineXL v3这个模型（kohaku delta也能，但是效果稍弱）本身就能使用提示词直出plana和arona，这个lokr旨在通过训练进一步还原这两个人物的人物特征；

• 初步测试未观察到明显影响画风，可考虑与画风/艺术家提示词或者是画风lora一起使用；

• 未测试这两个版本的lokr在非对应底模的迁移性。

训练参数

遵循Fair AI Public License 1.0-SD协议,公开训练参数,在kohya-ss的训练脚本上进行的训练。

• for an3xl v1.0版本，训练集共647张图片。

pretrained_model_name = "animagineXLV3_v30.safetensors"
learning_rate = 6e-06
resolution = "1024,1024"
lr_scheduler = "cosine"
train_batch_size = "4"
gradient_accumulation_steps = 4
max_train_steps = "1200"
mixed_precision = "fp16"
seed = "1919810"
optimizer_type = "lion8bit"
network_module = lycoris.kohya
network_args "preset=attn-mlp" "factor=8" "algo=lokr"
network_dim = 100000
multires_noise_iterations = "8" 
multires_noise_discount = "0.6" 
network_train_unet_only = true
min_snr_gamma = 10
enable_bucket = true
bucket_reso_steps = 32
min_bucket_reso = 512
max_bucket_reso = 2048
class_tokens = 'plana\(blue archive\)'
shuffle_caption = true
gradient_checkpointing = true
xformers = true
cache_latents_to_disk = true
persistent_data_loader_workers = true

• for delta v1.0版本，训练集共722张图片。

pretrained_model_name = "kohakuXLDelta_rev1.safetensors"
learning_rate = 1e-05
resolution = "1024,1024"
lr_scheduler = "cosine"
train_batch_size = "4"
gradient_accumulation_steps = 4
max_train_steps = "1200"
mixed_precision = "fp16"
seed = "1919810"
optimizer_type = "lion8bit"
network_module = lycoris.kohya
network_args "preset=attn-mlp" "factor=8" "algo=lokr"
network_dim = 100000
multires_noise_iterations = "8" 
multires_noise_discount = "0.6" 
network_train_unet_only = true
min_snr_gamma = 10
enable_bucket = true
bucket_reso_steps = 32
min_bucket_reso = 512
max_bucket_reso = 2048
class_tokens = 'plana\(blue archive\)'
shuffle_caption = true
gradient_checkpointing = true
xformers = true
cache_latents_to_disk = true
persistent_data_loader_workers = true

执照

Fair AI Public License 1.0-SD