Pixiv：https://www.pixiv.net/artworks/109004467
pretrained_model_name_or_path = "./sd-models/model.ckpt" v2 = false train_data_dir = "./train/guizhengchao" prior_loss_weight = 1 resolution = "1280,1280" enable_bucket = true min_bucket_reso = 256 max_bucket_reso = 1_536 output_name = "guizhengchao" output_dir = "./output" save_model_as = "safetensors" save_precision = "fp16" save_every_n_epochs = 1 max_train_epochs = 10 train_batch_size = 2 gradient_checkpointing = false network_train_unet_only = false network_train_text_encoder_only = false learning_rate = 0.00007 unet_lr = 0.00007 text_encoder_lr = 0.000008 lr_scheduler = "cosine_with_restarts" optimizer_type = "Lion" min_snr_gamma = 5 lr_scheduler_num_cycles = 3 network_module = "lycoris.kohya" network_dim = 16 network_alpha = 8 network_dropout = 0 log_with = "tensorboard" logging_dir = "./logs" caption_extension = ".txt" shuffle_caption = true weighted_captions = false keep_tokens = 0 max_token_length = 255 multires_noise_iterations = 6 multires_noise_discount = 0.3 seed = 1_337 clip_skip = 2 mixed_precision = "fp16" xformers = true lowram = false cache_latents = true cache_latents_to_disk = false persistent_data_loader_workers = true network_args = [ "conv_dim=5", "conv_alpha=1", "dropout=0", "algo=loha" ]
(((v1.0))) 推荐0.5到0.8
NAi上训练 7e-5/7e-5/8e-6 1280x1280分辨率 63张图 6000多步 epoch10 batchsize2 三个概念文件夹 分别是others /white pantyhose /black pantyhose
图片训练集为guizhencao.zip，可以下载使用
训练出的模型为guizhengcao.safetensors
已经上传10个epoch，我比较推荐第5个也就是guizhengchao-000005.safetensors，如果可以你也可以试试看其他的epoch，（就当顺便帮我测试一下模型(^_^)）
(((v1.0))) 0.5 to 0.8 is recommended
Training on NAi 7e-5/7e-5/8e-6 1280x1280 resolution 63 pictures More than 6000 steps epoch10 batchsize2 Three concept folders are others /white pantyhose /black pantyhose
The picture training set is guizhencao.zip, which can be downloaded and used
The trained model is guizhengcao.safetensors
10 epochs have been uploaded. I recommend the 10th one, which is guizhengcao.safetensors. If you can, you can also try other epochs (just help me test the model by the way (^_^))