AGM Style AKA 阿戈魔agm (Omone Hokoma AGM)
详情
下载文件
模型描述
触发词非必需，但
illustration_by_omone_hokoma_agm
已被添加到每个标题中，似乎确实增强了效果，但往往会大量添加AGM的签名风格，因此我建议：如果你使用触发词，可在负面提示中加入“signature,watermark,artist_name”以移除它。不过我会在手动使用Photoshop逐一删除所有数据集图像中的签名并基于此重新训练后更新这一建议，届时将无需额外的负面词。
以下是参数设置
训练信息▼
标题中最常见的标签
illustration_by_omone_hokoma_agm
100%
数据集文件夹结构
名称
40_agm
图像数量
595
重复次数
40
总图像数
23800
（总计）
595
图像数量
595
重复次数
40
总图像数
23800
训练参数
{
"ss_adaptive_noise_scale": "None",
"ss_caption_dropout_rate": "0.0",
"ss_steps": "8500",
"ss_noise_offset": "None",
"ss_sd_scripts_commit_hash": "15dd0a638af86f89dd0c457428e165598d4884a2",
"ss_num_batches_per_epoch": "23800",
"ss_color_aug": "False",
"ss_epoch": "0",
"ss_total_batch_size": "1",
"ss_network_alpha": "64.0",
"ss_ip_noise_gamma": "None",
"ss_num_epochs": "1",
"ss_session_id": "3173715531",
"ss_network_dim": "64",
"ss_keep_tokens": "0",
"ss_learning_rate": "1.0",
"ss_new_sd_model_hash": "e6bb9ea85bbf7bf6478a7c6d18b71246f22e95d41bcdd80ed40aa212c33cfeff",
"ss_lr_warmup_steps": "0",
"ss_optimizer": "prodigyopt.prodigy.Prodigy",
"ss_caption_dropout_every_n_epochs": "0",
"ss_network_module": "lycoris.kohya",
"ss_reg_dataset_dirs": "{}",
"ss_sd_model_hash": "be9edd61",
"ss_gradient_accumulation_steps": "1",
"ss_bucket_no_upscale": "False",
"ss_bucket_info": "null",
"ss_full_fp16": "False",
"ss_mixed_precision": "fp16",
"ss_network_dropout": "0.0",
"ss_gradient_checkpointing": "True",
"ss_random_crop": "False",
"ss_prior_loss_weight": "1.0",
"ss_max_grad_norm": "1.0",
"ss_max_bucket_reso": "None",
"ss_training_comment": "None",
"ss_num_reg_images": "0",
"ss_max_train_steps": "10000",
"ss_min_snr_gamma": "10.0",
"ss_num_train_images": "23800",
"ss_network_args": "{\"conv_dim\": \"64\", \"conv_alpha\": \"64\", \"factor\": \"-1\", \"use_cp\": \"True\", \"algo\": \"lokr\", \"dropout\": 0.0}",
"ss_shuffle_caption": "False",
"ss_unet_lr": "1.0",
"ss_resolution": "(1024, 1024)",
"ss_batch_size_per_device": "1",
"ss_multires_noise_discount": "0.2",
"ss_flip_aug": "False",
"ss_text_encoder_lr": "1.0",
"ss_lr_scheduler": "constant",
"ss_min_bucket_reso": "None",
"ss_zero_terminal_snr": "False",
"ss_lowram": "False",
"ss_seed": "12345",
"ss_multires_noise_iterations": "6",
"ss_base_model_version": "sdxl_base_v1-0",
"ss_enable_bucket": "False",
"ss_training_started_at": "1708377854.7039077",
"ss_clip_skip": "None",
"ss_v2": "False",
"ss_caption_tag_dropout_rate": "0.0",
"ss_max_token_length": "225",
"ss_output_name": "AGMXL",
"ss_scale_weight_norms": "None",
"ss_training_finished_at": "1708441002.8566256",
"ss_sd_model_name": "sdXL_v10VAEFix.safetensors",
"ss_cache_latents": "True",
"ss_face_crop_aug_range": "None"
}
这是我的第二个LoRA，第一个是角色LoRA，因此我想尝试一个风格LoRA，目前尚未见到类似SD1.5中那样针对AGM风格的SDXL LoRA。和我的第一个LoRA一样，我上传了我实际使用的完整训练数据，希望更有经验的用户能基于此数据进一步优化。我为每张图像都添加了标题，但在测试中发现其影响不大，因此无需触发词。如果此模型对你有效，请留下评分和评论，反馈对我非常重要，谢谢！