Yuzuriha (Hell's Paradise)
详情
下载文件
模型描述
我想听听其他LoRA创作者的意见，如果你有时间，我们可以聊一聊。
如果你使用过这个LoRA，请留下评价，无论好坏，我都想知道它是否对你有效，或者你是否遇到任何问题——反馈对我来说非常重要，谢谢！
这是我第一个LoRA，可能不是最好的，但我已在多个动漫XL模型上测试过，例如：
我发现，当LoRA权重设置在0.4到0.6之间时效果最好；如果想让角色穿上普通服装，0.5到0.7似乎比较合适；而如果你使用日本服饰，甚至可以调得更低。我使用Kohya_SS GUI训练了这个LoRA，由于我不清楚哪些设置最适合角色LoRA，所以我选择了其中一个XL预设。我也会在下方贴出这些设置，并上传我使用的完整数据集（包含标注文件），也许有比我更有经验的人能在此基础上进一步优化这个LoRA。
{ "ss_training_finished_at": "1705714270.261762", "ss_base_model_version": "sdxl_base_v1-0", "ss_network_dim": "Dynamic", "ss_sd_model_name": "stabilityai/stable-diffusion-xl-base-1.0", "ss_network_dropout": "None", "ss_reg_dataset_dirs": "{}", "ss_caption_tag_dropout_rate": "0.0", "ss_prior_loss_weight": "1.0", "ss_lowram": "False", "ss_adaptive_noise_scale": "None", "ss_mixed_precision": "bf16", "ss_color_aug": "False", "ss_num_epochs": "25", "ss_clip_skip": "None", "ss_epoch": "4", "ss_ip_noise_gamma": "None", "ss_total_batch_size": "6", "ss_max_bucket_reso": "2048", "ss_output_name": "yuzuriha_\(hell's_paradise\)", "ss_bucket_no_upscale": "True", "ss_unet_lr": "1.0", "ss_sd_scripts_commit_hash": "7dd2a273b225fb33f8a8769b1af77517dd95d3a9", "ss_max_grad_norm": "1.0", "ss_scale_weight_norms": "None", "ss_full_fp16": "False", "ss_seed": "1337", "ss_lr_warmup_steps": "533", "ss_cache_latents": "True", "ss_noise_offset": "0.0357", "ss_steps": "1750", "ss_multires_noise_iterations": "None", "ss_max_train_steps": "10667", "ss_min_snr_gamma": "5.0", "ss_debiased_estimation": "False", "ss_multires_noise_discount": "0.3", "ss_caption_dropout_rate": "0.0", "ss_face_crop_aug_range": "None", "ss_caption_dropout_every_n_epochs": "0", "ss_training_started_at": "1705698819.5353682", "ss_flip_aug": "False", "ss_max_token_length": "225", "ss_training_comment": "Dynamic resize with sv_fro: 0.9 from 256; None", "ss_num_batches_per_epoch": "427", "ss_random_crop": "False", "ss_shuffle_caption": "False", "ss_num_train_images": "2560", "ss_batch_size_per_device": "6", "ss_session_id": "3385709643", "ss_enable_bucket": "True", "ss_optimizer": "dadaptation.dadapt_adam.DAdaptAdam(decouple=True,weight_decay=0.2,betas=(0.9, 0.99),growth_rate=1.02)", "ss_min_bucket_reso": "256", "ss_gradient_checkpointing": "True", "ss_network_alpha": "Dynamic", "ss_learning_rate": "1.0", "ss_bucket_info": "{"buckets": {"0": {"resolution": [1024, 1024], "count": 2560}}, "mean_img_ar_error": 0.0}", "ss_text_encoder_lr": "1.0", "ss_keep_tokens": "0", "ss_resolution": "(1024, 1024)", "ss_v2": "False", "ss_gradient_accumulation_steps": "1", "ss_lr_scheduler": "cosine", "ss_num_reg_images": "0", "ss_zero_terminal_snr": "False", "ss_network_module": "networks.lora" }
不过我注意到，这个LoRA的体积高达1.7GB，对于一个角色LoRA来说太大了。于是我使用Kohya_SS GUI的LoRA工具，将256维降为32维，前后测试效果几乎一致，最终文件大小从近2GB缩减至170MB。