Software lllyasviel/Webui-forge
trained picture prompt
【3D】【guofeng】【illustration】【toon】【tangfu,chinese traditional hanfu,qixiong ruqun】【photo-referenced】
Reg picture prompt
【1photo】【1woman32】【7788RE】【7788WH】【ALAN】【FENGJING】【HAILUOKE】【LINGXIAOQIAN】【NIGHTYY】【XIAOZHUANGFOX】【travel】
----训练图片提示词与正则化提示词
This LORA model is primarily trained on realistic images.
use【realistic】
----这是个以真实为主要训练的LORA 【realistic】是它的主要训练风格
There are only 200+ training images and 3,100 regularized images,Realistic style accounts for 70% training images, foundational model is NOOBAI-eps1.1 .
----训练图片只有200+张与3100+正则化图片，真实化风格图片占训练图片的70%，基础模型是noobai-eps1.1
However, the output image bears some resemblance to UEngine or Guofeng2 in SD1.5. if want more photographing, it should be use【photo-referenced】.
----但是输出的图片有点像虚幻引擎或者sd1.5的guofeng2 .照片化就加
【photo-referenced】
In hires.fix, avoid using Karras noise, as it may lead to excessive noise or even overexposure.
----在hires.fix 别用Karras 噪声会多甚至过曝.
The <style> of the website linked below in is flexible.
----与下面<STYLE>网址的风格可变,
It doesn't have to be 【realistic】.
是除了真实风格以外其他风格均可适配.
<STYLE>
Description
Deleted some regularization prompts,
model_train_type = "sdxl-lora"
pretrained_model_name_or_path = noobaiXLNAIXL_epsilonPred11Version.safetensors
train_data_dir = #######################
reg_data_dir = #######################
prior_loss_weight = 0.7
resolution = "1248,1248"
enable_bucket = true
min_bucket_reso = 1024
max_bucket_reso = 3840
bucket_reso_steps = 64
bucket_no_upscale = true
output_name = #######################
output_dir = #######################
save_model_as = "safetensors"
save_precision = "bf16"
save_every_n_epochs = 1
save_state = true
max_train_epochs = 20
train_batch_size = 8
gradient_checkpointing = true
gradient_accumulation_steps = 2
network_train_unet_only = false
network_train_text_encoder_only = false
learning_rate = 0.00016
unet_lr = 0.00016
text_encoder_lr = 0.00016
lr_scheduler = "constant"
lr_warmup_steps = 0
optimizer_type = "AdamW8bit"
network_module = "networks.lora"
network_dim = 96
network_alpha = 48
log_with = "tensorboard"
logging_dir = "./logs"
caption_extension = ".txt"
shuffle_caption = false
keep_tokens = 0
max_token_length = 400
noise_offset = 0.08
color_aug = false
flip_aug = false
random_crop = false
seed = 1337
mixed_precision = "bf16"
full_bf16 = true
xformers = true
lowram = false
cache_latents = true
cache_latents_to_disk = true
persistent_data_loader_workers = true
vae_batch_size = 8
ddp_gradient_as_bucket_view = false
no_metadata = false
lr_scheduler_num_cycles = 1
network_args = [
"rank_dropout=0.25",
"module_dropout=0.15",
"conv_dim=64",
"conv_alpha=32"
]
USE 30.5G VRAM