配置如下model_train_type = "sdxl-lora"
pretrained_model_name_or_path = "/root/autodl-tmp/ill/illustriousXL_v01.safetensors etensors,illustriousXL_v01.safetensors) etensors,illustriousXL_v01.safetensors) "
resume = ""
train_data_dir = "/root/autodl-tmp/lora-scripts/train/theresa/"
prior_loss_weight = 1
resolution = "1024,1024"
enable_bucket = true
min_bucket_reso = 1
max_bucket_reso = 2560
bucket_reso_steps = 32
bucket_no_upscale = true
output_name = "theresav2"
output_dir = "/root/autodl-tmp/lora-scripts/output/theresav2"
save_model_as = "safetensors"
save_precision = "fp16"
save_every_n_epochs = 2
save_state = true
max_train_epochs = 50
train_batch_size = 4
gradient_checkpointing = true
gradient_accumulation_steps = 1
network_train_unet_only = true
network_train_text_encoder_only = false
learning_rate = 1
unet_lr = 1
text_encoder_lr = 1
lr_scheduler = "cosine"
lr_warmup_steps = 0
optimizer_type = "Prodigy"
network_module = "networks.lora"
network_dim = 32
network_alpha = 32
network_dropout = 0.01
sample_prompts = "masterpiece, best quality, theresa,1girl ,full body,turtleneck dress,high heel boots --w 1024 --h 1024 --l 7 --s 24 --d 1337"
sample_sampler = "euler_a"
sample_every_n_epochs = 2
log_with = "tensorboard"
log_prefix = "theresav2"
log_tracker_name = "LORA"
logging_dir = "./logs"
caption_extension = ".txt"
shuffle_caption = true
keep_tokens = 1
max_token_length = 255
seed = 1019
mixed_precision = "fp16"
full_fp16 = true
no_half_vae = true
xformers = true
sdpa = true
lowram = false
cache_latents = true
cache_latents_to_disk = true
persistent_data_loader_workers = true
optimizer_args = [
"decouple=True",
"weight_decay=0.01",
"use_bias_correction=True",
"d_coef=2.0"
]

prodigy的d*lr才是真正的学习率，它是冷启动的，最开始从0开始一点点自动上升，因此最开始训练的几轮看起来和原图变化不大。aw8bit如果不设置预热步数的话一开始学习率就是你设置的值，随后逐渐减少到0，所以一开始就有很大的效果，越到后面变化越小

我记得选用prodigy以后下面会有一个参数prodigy_d_coef，这个参数你填上2.0试试。还有看图上的黑色线显示只训练了两个epoch，正在训练epoch3，一共只有300多step，这时候lr还是接近0也是正常的，可能大概第6，7个epoch才能开始看到上涨。你可以把鼠标放在d*lr那个图上看看，曲线上的value值是0还是一个接近0的很小的数比如1e-8之类的

跑了十几轮一点变化都没有model_train_type = "sdxl-lora"
pretrained_model_name_or_path = "/root/autodl-tmp/ill/网页链接 "
resume = ""
train_data_dir = "/root/autodl-tmp/lora-scripts/train/theresa/"
prior_loss_weight = 1
resolution = "512,512"
enable_bucket = true
min_bucket_reso = 1
max_bucket_reso = 1024
bucket_reso_steps = 32
bucket_no_upscale = true
output_name = "theresav3"
output_dir = "/root/autodl-tmp/lora-scripts/output/theresav2"
save_model_as = "safetensors"
save_precision = "fp16"
save_every_n_epochs = 2
save_state = true
max_train_epochs = 40
train_batch_size = 4
gradient_checkpointing = true
gradient_accumulation_steps = 1
network_train_unet_only = false
network_train_text_encoder_only = false
learning_rate = 1
unet_lr = 1
text_encoder_lr = 1
lr_scheduler = "constant"
lr_warmup_steps = 0
optimizer_type = "Prodigy"
network_module = "网页链接 "
network_dim = 32
network_alpha = 32
network_dropout = 0.01
sample_prompts = "masterpiece, best quality, 1girl, solo,theresa,turtleneck dress, chest-shoulder device,single pauldron, high heel boots,sky,grass --w 1024 --h 1024 --l 7 --s 24 --d 1337"
sample_sampler = "euler_a"
sample_every_n_epochs = 1
log_with = "tensorboard"
log_prefix = "theresav1"
log_tracker_name = "LORA"
logging_dir = "./logs"
caption_extension = ".txt"
shuffle_caption = true
keep_tokens = 1
max_token_length = 255
seed = 1019
mixed_precision = "fp16"
full_fp16 = true
no_half_vae = true
xformers = true
sdpa = true
lowram = false
cache_latents = true
cache_latents_to_disk = true
persistent_data_loader_workers = true
optimizer_args = [
"decouple=True",
"weight_decay=0.01",
"use_bias_correction=True",
"d_coef=2.0"
]

我学的是flux lora trainer，只要跑过程loss曲线慢慢下降的模型 最终效果都还不错。constant的意思是固定学习率 所以你跑的lr是直线。我用的adafactor优化器。 给你做个参考。