点滴 intravenous drip-XLPony
ファイルをダウンロード
このバージョンについて
点滴 v001
コマンドライン設定
accelerate.EXE launch --dynamo_backend no --dynamo_mode default --mixed_precision bf16 --num_processes 2 --num_machines 1 --num_cpu_threads_per_process 2sdxl_train_network.py --config_file config_lora-20240618-184456.toml--lr_scheduler_type CosineAnnealingLR--lr_scheduler_args T_max=1040 eta_min=1e-7--network_args loraplus_lr_ratio=16--alpha_mask
toml設定ファイル
bucket_no_upscale = truebucket_reso_steps = 64cache_latents = truecache_latents_to_disk = truecaption_dropout_rate = 0.05caption_extension = ".txt"clip_skip = 1dynamo_backend = "no"enable_bucket = trueepoch = 100gradient_accumulation_steps = 1gradient_checkpointing = truehuber_c = 0.1huber_schedule = "snr"keep_tokens = 1learning_rate = 0.0001log_with = "wandb"logging_dir = "log"loss_type = "smooth_l1"lr_scheduler = "cosine"lr_scheduler_args = []lr_scheduler_num_cycles = 1lr_scheduler_power = 1max_bucket_reso = 2048max_data_loader_n_workers = 0max_grad_norm = 0.1max_timestep = 1000max_token_length = 225max_train_epochs = 100max_train_steps = 10400mem_eff_attn = truemin_bucket_reso = 256min_snr_gamma = 5mixed_precision = "bf16"network_alpha = 32network_args = ["rank_dropout=0.2", "module_dropout=0.2"]network_dim = 32network_dropout = 0.2network_module = "networks.lora"no_half_vae = truenoise_offset_type = "Original"optimizer_args = ["weight_decay=0.01"]optimizer_type = "AdamW8bit"output_dir = "output"output_name = "intravenousdrip-XLPony_001"pretrained_model_name_or_path = "ponyDiffusionV6XL_v6StartWithThisOne.safetensors"prior_loss_weight = 0.1resolution = "1024,1024"sample_prompts = "prompt.txt"sample_sampler = "euler_a"save_every_n_epochs = 10save_model_as = "safetensors"save_precision = "bf16"save_state = trueseed = 1shuffle_caption = truetext_encoder_lr = 1e-5train_batch_size = 1unet_lr = 0.0001xformers = true
モデル説明
いつも通りdanbooruやreluなどから画像を集めたら、エロ画像のデータが異常に増えた。
治療中の画像などは、比較的センシティブな個人情報であるため、収集が難しいのも納得できる。
v002を作るにあたり、良い画像のおすすめがあれば教えてください。
v001
タグの手動整備なし
10400step 100ep


