1. 安裝環境 #

“LoRA Easy Training Scripts"這個Python程式Linux和Windows都可以用，下面以Ubuntu為例。

安裝 Anaconda，建立虛擬環境

conda create --name loratraining python=3.10.6
conda activate loratraining

複製儲存庫

git clone https://github.com/derrian-distro/LoRA_Easy_Training_Scripts.git
cd LoRA_Easy_Training_Scripts
git submodule init
git submodule update
cd sd_scripts
pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu116
pip install --upgrade -r requirements.txt
pip install -U xformers

設定加速選項

accelerate config
#依序回答：
#- This machine
#- No distributed training
#- NO
#- NO
#- NO
#- all
#- fp16

LoRA的訓練資料目錄結構不太一樣，需建立目錄結構如下。已經上好提示詞的訓練資料要放在img_dir下面，將目錄名稱取名為數字_概念，目錄名稱前面加上數字代表要重複的步數。

新增訓練設定檔trainingconfig.json

vim trainingconfig.json

填入以下內容(雙斜線的註解記得刪除) LoRA的總訓練步數計算公式為：訓練圖片數量 × 重複次數 ÷ train_batch_size × epoch

{
  //基於何種模型訓練
  "pretrained_model_name_or_path": "/home/user/桌面/heralora/anything-v4.5-pruned.ckpt",
  "v2": false,
  "v_parameterization": false,
  //紀錄檔輸出目錄
  "logging_dir": "/home/user/桌面/heralora/log_dir/",
  //訓練資料目錄
  "train_data_dir": "/home/user/桌面/heralora/image_dir/",
  //註冊目錄
  "reg_data_dir": "/home/user/桌面/heralora/reg_dir/",
  //輸出目錄
  "output_dir": "/home/user/桌面/heralora/output_dir",
  //訓練的圖片最大長寬
  "max_resolution": "512,512",
  //學習率
  "learning_rate": "1e-5",
  "lr_scheduler": "constant_with_warmup",
  "lr_warmup": "5",
  "train_batch_size": 3,
  //訓練時期
  "epoch": "4",
  "save_every_n_epochs": "",
  "mixed_precision": "fp16",
  "save_precision": "fp16",
  "seed": "",
  "num_cpu_threads_per_process": 32,
  "cache_latents": true,
  "caption_extension": ".txt",
  "enable_bucket": true,
  "gradient_checkpointing": false,
  "full_fp16": false,
  "no_token_padding": false,
  "stop_text_encoder_training": 0,
  "use_8bit_adam": true,
  "xformers": true,
  "save_model_as": "safetensors",
  "shuffle_caption": true,
  "save_state": false,
  "resume": "",
  "prior_loss_weight": 1.0,
  "text_encoder_lr": "1.5e-5",
  "unet_lr": "1.5e-4",
  "network_dim": 128,
  "lora_network_weights": "",
  "color_aug": false,
  "flip_aug": false,
  "clip_skip": 2,
  "mem_eff_attn": false,
  "output_name": "",
  "model_list": "",
  "max_token_length": "150",
  "max_train_epochs": "",
  "max_data_loader_n_workers": "",
  "network_alpha": 128,
  "training_comment": "",
  "keep_tokens": 2,
  "lr_scheduler_num_cycles": "",
  "lr_scheduler_power": "",
  "persistent_data_loader_workers": true,
  "bucket_no_upscale": true,
  "random_crop": false,
  "caption_dropout_every_n_epochs": 0.0,
  "caption_dropout_rate": 0
}

2. 開始訓練 #

有些系統需要指定CUDA安裝路徑

export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

輸入以下指令，載入json設定檔。libnvinfer.so.7: cannot open shared object file的警告可以暫時忽略。

accelerate launch main.py --load_json_path "/home/user/trainingconfig.json"

之後會自動開始訓練。訓練好的模型位於訓練設定檔所寫的output_dir目錄。將.safetensors檔移動至SD WebUI根目錄下的/models/Lora。

3. LoRA模型使用方式 #

點選SD WebUI右上角，Show extra networks

點選要使用的LoRA，將其加入至提示詞欄位

再加上訓練時使用的提示詞，即可生成使用LoRA風格的人物。