<h3 id="heading-33">更新</h3>基于Pony模型训练了一版人物lycoris，可适用于pony系模型。参数见版本相关。<h3 id="heading-1874">导言</h3>这是一个《Umamusume Pretty Deby》<s>《闪耀</s><s>优俊少女</s><s>》</s>的初期企划大震撼角色LyCORIS。<h3 id="heading-1875">使用方法</h3>建议以0.8左右的权重使用。在打标过程中，为隔离元素污染，采用了每个卡面分开设关键词的方法，以下是各个立绘的关键词：特雷森制服：tracen_unifrom决胜服：shobofuku<h3 id="heading-1876">训练细节</h3>以NAI Final作为底模训练，使用了约36张图片作为训练集，生成同等数量的类似图片作为回归训练集，训练20个epoch，最终步数为3600步。以下为主要设置参数：network_dim=32network_alpha=32keep_tokens=1conv_dim=4conv_alpha=4lr="1e-5"unet_lr="1e-5"text_encoder_lr="1e-5"batch_size = 2optimizer = PagedAdamW8bits<h3 id="heading-1877">局限性</h3><ul><li>人物立绘中的不对称设计的还原难度非常大，基本上只能依靠随机抽卡和手动辅助达到一定的还原效果。</li><li>图像容易过饱和，建议在此类情况发生时降低权重或CFG</li></ul><h3 id="heading-1878">Introduction</h3>This is a LyCORIS of Deep Impact from "Umamusum Pretty Deby".<h3 id="heading-1879">Usage</h3>Suggest using a weight of around 0.75.During the marking process, in order to isolate element contamination, a method of setting keywords separately for each card surface was adopted. The following are the keywords for each vertical drawing:Teresan uniform: tracen_unifromShobofuku：shobofuku<h3 id="heading-1880">Training Details</h3>Using NAI Final as the base model for training, approximately 36 images were used as the training set, and about half of the images were used as the regression training set. Final step count of 3600.The following are the main settings parameters:network_ dim=32network_ alpha=32keep_ tokens=1conv_ dim=4conv_ alpha=4lr="1e-5"unet_ lr="1e-5"text_ encoder_ lr="1e-5"batch_ size = 2optimizer = PagedAdamW8bits<h3 id="heading-1881">Limitations</h3><ol><li>It is very hard to draw the asymmetrical designs in character portraits. Maybe a lot of tries and manual assistance is a certain way to achieve good results.</li><li>Images are prone to oversaturation, and it is recommended to reduce weight or CFG when such situations occur</li></ol>

Deep Impact - Umamusume Pretty Deby

更新基于Pony模型训练了一版人物lycoris，可适用于pony系模型。参数见下：UpdateA version of the character Lycoris was trained based on the Pony model, which can be applied to the Pony series model.The parameters are as follows:{
 "LoRA_type": "LyCORIS/LoCon",
 "LyCORIS_preset": "full",
 "adaptive_noise_scale": 0,
 "additional_parameters": "",
 "block_alphas": "",
 "block_dims": "",
 "block_lr_zero_threshold": "",
 "bucket_no_upscale": true,
 "bucket_reso_steps": 128,
 "cache_latents": false,
 "cache_latents_to_disk": false,
 "caption_dropout_every_n_epochs": 0.0,
 "caption_dropout_rate": 0,
 "caption_extension": ".txt",
 "clip_skip": 2,
 "color_aug": false,
 "constrain": 0.0,
 "conv_alpha": 4,
 "conv_block_alphas": "",
 "conv_block_dims": "",
 "conv_dim": 8,
 "debiased_estimation_loss": false,
 "decompose_both": false,
 "dim_from_weights": false,
 "down_lr_weight": "",
 "enable_bucket": true,
 "epoch": 10,
 "factor": -1,
 "flip_aug": false,
 "fp8_base": false,
 "full_bf16": false,
 "full_fp16": false,
 "gpu_ids": "",
 "gradient_accumulation_steps": "1",
 "gradient_checkpointing": true,
 "keep_tokens": 1,
 "learning_rate": 1.0,
 "log_tracker_config": "",
 "log_tracker_name": "",
 "logging_dir": "",
 "lora_network_weights": "",
 "lr_scheduler": "cosine",
 "lr_scheduler_args": "",
 "lr_scheduler_num_cycles": "",
 "lr_scheduler_power": "",
 "lr_warmup": 0,
 "max_bucket_reso": 4096,
 "max_data_loader_n_workers": "0",
 "max_grad_norm": 1,
 "max_resolution": "1024,1024",
 "max_timestep": 1000,
 "max_token_length": "225",
 "max_train_epochs": "",
 "max_train_steps": "",
 "mem_eff_attn": false,
 "mid_lr_weight": "",
 "min_bucket_reso": 1024,
 "min_snr_gamma": 0,
 "min_timestep": 0,
 "mixed_precision": "bf16",
 "model_list": "custom",
 "module_dropout": 0,
 "multi_gpu": false,
 "multires_noise_discount": 0.1,
 "multires_noise_iterations": 6,
 "network_alpha": 8,
 "network_dim": 16,
 "network_dropout": 0,
 "noise_offset": 0,
 "noise_offset_type": "Multires",
 "num_cpu_threads_per_process": 2,
 "num_machines": 1,
 "num_processes": 1,
 "optimizer": "Prodigy",
 "optimizer_args": "weight_decay=0.01 d_coef=1 use_bias_correction=True safeguard_warmup=False betas=0.9,0.99 decouple=True ",
 "output_dir": "/root/autodl-tmp/output",
 "output_name": "deep_impact-ponyXL",
 "persistent_data_loader_workers": false,
 "pretrained_model_name_or_path": "/root/kohya_ss/models/ponyDiffusionV6XL.safetensors",
 "prior_loss_weight": 1.0,
 "random_crop": false,
 "rank_dropout": 0,
 "rank_dropout_scale": false,
 "reg_data_dir": "/root/autodl-tmp/reg_deep_impact__XL",
 "rescaled": false,
 "resume": "",
 "sample_every_n_epochs": 1,
 "sample_every_n_steps": 0,
 "sample_prompts": "score_9, score_8_up, score_7_up, anime BREAK, shobofuku, 1girl, solo, long hair, looking at viewer, skirt, simple background, brown hair, hair ornament, thighhighs, gloves, long sleeves, white background, animal ears, closed mouth, purple eyes, tail, full body, boots, white gloves, signature, cape, official alternate costume, hand on hip, glowing, horse ears, horse girl, glowing eyes, horse tail, glowing eye --n source_pony, source_furry, source_cartoon, greyscale, from below --w 1024 --h 1024\n ",
 "sample_sampler": "euler_a",
 "save_every_n_epochs": 1,
 "save_every_n_steps": 0,
 "save_last_n_steps": 0,
 "save_last_n_steps_state": 0,
 "save_model_as": "safetensors",
 "save_precision": "bf16",
 "save_state": false,
 "scale_v_pred_loss_like_noise_pred": false,
 "scale_weight_norms": 0,
 "sdxl": true,
 "sdxl_cache_text_encoder_outputs": false,
 "sdxl_no_half_vae": true,
 "seed": "31337",
 "shuffle_caption": true,
 "stop_text_encoder_training_pct": 0,
 "text_encoder_lr": 1.0,
 "train_batch_size": 2,
 "train_data_dir": "/root/autodl-tmp/deep_impact__XL",
 "train_norm": false,
 "train_on_input": true,
 "training_comment": "",
 "unet_lr": 1.0,
 "unit": 1,
 "up_lr_weight": "",
 "use_cp": false,
 "use_scalar": false,
 "use_tucker": false,
 "use_wandb": false,
 "v2": false,
 "v_parameterization": false,
 "v_pred_like_loss": 0,
 "vae": "/root/kohya_ss/models/sdxl_vae_fp16fix.safetensors",
 "vae_batch_size": 0,
 "wandb_api_key": "",
 "wandb_run_name": "",
 "weighted_captions": false,
 "xformers": "xformers"
}<pre><code></code></pre>

deep_impact-ponyXL.safetensors

Pony

v1.0

更新基于Pony模型训练了一版人物lycoris，可适用于pony系模型。参数见版本相关。导言这是一个《Umamusume Pretty Deby》<s>《闪耀</s><s>优俊少女</s><s>》</s>的初期企划大震撼角色LyCORIS。使用方法建议以0.8左右的权重使用。在打标过程中，为隔离元素污染，采用了每个卡面分开设关键词的方法，以下是各个立绘的关键词：特雷森制服：tracen_unifrom决胜服：shobofuku训练细节以NAI Final作为底模训练，使用了约36张图片作为训练集，生成同等数量的类似图片作为回归训练集，训练20个epoch，最终步数为3600步。以下为主要设置参数：network_dim=32network_alpha=32keep_tokens=1conv_dim=4conv_alpha=4lr="1e-5"unet_lr="1e-5"text_encoder_lr="1e-5"batch_size = 2optimizer = PagedAdamW8bits局限性<ul><li>人物立绘中的不对称设计的还原难度非常大，基本上只能依靠随机抽卡和手动辅助达到一定的还原效果。</li><li>图像容易过饱和，建议在此类情况发生时降低权重或CFG</li></ul>IntroductionThis is a LyCORIS of Deep Impact from "Umamusum Pretty Deby".UsageSuggest using a weight of around 0.75.During the marking process, in order to isolate element contamination, a method of setting keywords separately for each card surface was adopted. The following are the keywords for each vertical drawing:Teresan uniform: tracen_unifromShobofuku：shobofukuTraining DetailsUsing NAI Final as the base model for training, approximately 36 images were used as the training set, and about half of the images were used as the regression training set. Final step count of 3600.The following are the main settings parameters:network_ dim=32network_ alpha=32keep_ tokens=1conv_ dim=4conv_ alpha=4lr="1e-5"unet_ lr="1e-5"text_ encoder_ lr="1e-5"batch_ size = 2optimizer = PagedAdamW8bitsLimitations<ol><li>It is very hard to draw the asymmetrical designs in character portraits. Maybe a lot of tries and manual assistance is a certain way to achieve good results.</li><li>Images are prone to oversaturation, and it is recommended to reduce weight or CFG when such situations occur</li></ol>

Files with SHA256: 21ea7b06bc2760e9f21f1e9c3f971bc43a71f16e7b7656a3a90d5fd7528fb7ee