chinese

chinese-hanfu-sd1.5-v30 训练日记

训练数据：
found directory /dataset/train_dataset2/chinese-hanfu-sd1-v30/img/10_ohxm woman contains 2465 image files
found directory /dataset/train_dataset2/chinese-hanfu-sd1-v30/img/10_khs woman contains 8220 image files
found directory /dataset/train_dataset2/chinese-hanfu-sd1-v30/img/10_sks woman contains 799 image files
114840 train images with repeating.

全部的数据

从SD1.5的基础模型开始训练。

accelerate launch --num_cpu_threads_per_process=2 "./train_db.py" \
--enable_bucket \
--min_bucket_reso=384  \
--max_bucket_reso=1280  \
--pretrained_model_name_or_path="/dataset/pretrain/v1-5-pruned-emaonly.safetensors" \
--train_data_dir="/dataset/train_dataset2/chinese-hanfu-sd1-v30/img" \
--resolution="768,768" \
--output_dir="/dataset/train_dataset2/chinese-hanfu-sd1-v30/model" \
--logging_dir="/dataset/train_dataset2/chinese-hanfu-sd1-v30/log" \
--save_model_as=safetensors \
--prior_loss_weight=1 \
--output_name="chinese-hanfu-sd1.5-v30" \
--lr_scheduler_num_cycles="20" \
--max_data_loader_n_workers="0" \
--learning_rate="5e-6" \
--lr_scheduler="constant" \
--train_batch_size="8" \
--save_every_n_epochs="1" \
--mixed_precision="bf16"  \
--save_precision="fp16" \
--seed="1254534" \
--caption_extension=".txt" \
--cache_latents \
--cache_latents_to_disk \
--optimizer_type="AdamW8bit" \
--max_train_epochs=15 \
--max_data_loader_n_workers="0" \
--clip_skip=2 \
--bucket_reso_steps=64 \
--xformers \
--bucket_no_upscale \
--noise_offset=0.0357 \
--keep_tokens=1 \
--gradient_checkpointing \
--save_every_n_steps="10000"