在 Kohya 使用多解析度訓練，在 datasets 的配置上可以採用以下規則：
| 解析度 | 說明 |
|---|---|
[1024, 1024] |
訓練內容可以是全部的訓練子集 |
[768x 768] |
剔除 訓練目標偏小的訓練子集 |
[512, 512] |
僅保留 訓練目標最大的訓練子集 |
同時，每一個 datasets 都可以配置不同的 batch_size 以及訓練集圖片的重複次數 num_repeats。
訓練集圖片重複次數的邏輯跟以往訓練的數據平衡概念差不多。
而執行的 batch_size 可以在最小解析度中開高一點（使用 2 ~ 4），對於小圖面內容的穩定性可以增加。
| [general] | |
| bucket_reso_steps = 32 | |
| bucket_no_upscale = true | |
| min_bucket_reso = 256 | |
| max_bucket_reso = 2048 | |
| enable_bucket = true | |
| color_aug = false | |
| flip_aug = false | |
| random_crop = false | |
| shuffle_caption = false | |
| caption_separator = "," | |
| enable_wildcard = false | |
| caption_extension = ".txt" | |
| caption_dropout_rate = 0.0 | |
| caption_tag_dropout_rate = 0.0 | |
| [[datasets]] | |
| batch_size = 1 | |
| resolution = [1024, 1024] | |
| network_multiplier = 1.0 | |
| keep_tokens = 1 | |
| [[datasets.subsets]] | |
| image_dir = "/training_image/4_close-up" | |
| num_repeats = 4 | |
| is_reg = false | |
| class_tokens = "girl close-up" | |
| [[datasets.subsets]] | |
| image_dir = "/training_image/protrait" | |
| num_repeats = 1 | |
| is_reg = false | |
| class_tokens = "girl protrait" | |
| [[datasets.subsets]] | |
| image_dir = "/training_image/full-body" | |
| num_repeats = 3 | |
| is_reg = false | |
| class_tokens = "girl full-body" | |
| [[datasets]] | |
| batch_size = 2 | |
| resolution = [768, 768] | |
| network_multiplier = 1.0 | |
| keep_tokens = 1 | |
| [[datasets.subsets]] | |
| image_dir = "/training_image/close-up" | |
| num_repeats = 4 | |
| is_reg = false | |
| class_tokens = "girl close-up" | |
| [[datasets.subsets]] | |
| image_dir = "/training_image/protrait" | |
| num_repeats = 1 | |
| is_reg = false | |
| class_tokens = "girl protrait" | |
| [[datasets]] | |
| batch_size = 4 | |
| resolution = [512,512] | |
| network_multiplier = 1.0 | |
| keep_tokens = 1 | |
| [[datasets.subsets]] | |
| image_dir = "/training_image/close-up" | |
| num_repeats = 8 | |
| is_reg = false | |
| class_tokens = "girl close-up" |
| { | |
| "LoRA_type": "Flux1", | |
| "LyCORIS_preset": "full", | |
| "adaptive_noise_scale": 0, | |
| "additional_parameters": "--log_prefix=flux --network_train_unet_only", | |
| "ae": "/data/models/vae/ae.safetensors", | |
| "apply_t5_attn_mask": false, | |
| "async_upload": false, | |
| "block_alphas": "", | |
| "block_dims": "", | |
| "block_lr_zero_threshold": "", | |
| "bucket_no_upscale": true, | |
| "bucket_reso_steps": 32, | |
| "bypass_mode": false, | |
| "cache_latents": true, | |
| "cache_latents_to_disk": true, | |
| "caption_dropout_every_n_epochs": 0, | |
| "caption_dropout_rate": 0, | |
| "caption_extension": ".txt", | |
| "clip_l": "/data/models/clip/clip_l.safetensors", | |
| "clip_skip": 1, | |
| "color_aug": false, | |
| "constrain": 0, | |
| "conv_alpha": 1, | |
| "conv_block_alphas": "", | |
| "conv_block_dims": "", | |
| "conv_dim": 1, | |
| "dataset_config": "/data/Train/datasets_config.toml", | |
| "debiased_estimation_loss": false, | |
| "decompose_both": false, | |
| "dim_from_weights": false, | |
| "discrete_flow_shift": 3.1582, | |
| "dora_wd": false, | |
| "down_lr_weight": "", | |
| "dynamo_backend": "eager", | |
| "dynamo_mode": "default", | |
| "dynamo_use_dynamic": false, | |
| "dynamo_use_fullgraph": false, | |
| "enable_bucket": true, | |
| "epoch": 8, | |
| "extra_accelerate_launch_args": "", | |
| "factor": -1, | |
| "flip_aug": false, | |
| "flux1_cache_text_encoder_outputs": true, | |
| "flux1_cache_text_encoder_outputs_to_disk": true, | |
| "flux1_checkbox": true, | |
| "fp8_base": true, | |
| "full_bf16": false, | |
| "full_fp16": false, | |
| "gpu_ids": "", | |
| "gradient_accumulation_steps": 1, | |
| "gradient_checkpointing": true, | |
| "guidance_scale": 1, | |
| "highvram": true, | |
| "huber_c": 0.1, | |
| "huber_schedule": "snr", | |
| "huggingface_path_in_repo": "", | |
| "huggingface_repo_id": "", | |
| "huggingface_repo_type": "", | |
| "huggingface_repo_visibility": "", | |
| "huggingface_token": "", | |
| "ip_noise_gamma": 0, | |
| "ip_noise_gamma_random_strength": false, | |
| "keep_tokens": 1, | |
| "learning_rate": 5e-05, | |
| "log_config": false, | |
| "log_tracker_config": "", | |
| "log_tracker_name": "", | |
| "log_with": "", | |
| "logging_dir": "/data/Train/log", | |
| "loraplus_lr_ratio": 0, | |
| "loraplus_text_encoder_lr_ratio": 0, | |
| "loraplus_unet_lr_ratio": 0, | |
| "loss_type": "l2", | |
| "lowvram": false, | |
| "lr_scheduler": "constant_with_warmup", | |
| "lr_scheduler_args": "", | |
| "lr_scheduler_num_cycles": 1, | |
| "lr_scheduler_power": 1, | |
| "lr_scheduler_type": "", | |
| "lr_warmup": 5, | |
| "main_process_port": 0, | |
| "masked_loss": false, | |
| "max_bucket_reso": 2048, | |
| "max_data_loader_n_workers": 8, | |
| "max_grad_norm": 1, | |
| "max_resolution": "1024,1024", | |
| "max_timestep": 1000, | |
| "max_token_length": 225, | |
| "max_train_epochs": 0, | |
| "max_train_steps": 4000, | |
| "mem_eff_attn": false, | |
| "mem_eff_save": false, | |
| "metadata_author": "", | |
| "metadata_description": "", | |
| "metadata_license": "", | |
| "metadata_tags": "", | |
| "metadata_title": "", | |
| "mid_lr_weight": "", | |
| "min_bucket_reso": 256, | |
| "min_snr_gamma": 5, | |
| "min_timestep": 0, | |
| "mixed_precision": "bf16", | |
| "model_list": "custom", | |
| "model_prediction_type": "raw", | |
| "module_dropout": 0, | |
| "multi_gpu": false, | |
| "multires_noise_discount": 0.3, | |
| "multires_noise_iterations": 0, | |
| "network_alpha": 16, | |
| "network_dim": 8, | |
| "network_dropout": 0, | |
| "network_weights": "", | |
| "noise_offset": 0.03, | |
| "noise_offset_random_strength": false, | |
| "noise_offset_type": "Original", | |
| "num_cpu_threads_per_process": 2, | |
| "num_machines": 1, | |
| "num_processes": 1, | |
| "optimizer": "AdamW8bit", | |
| "optimizer_args": "weight_decay=0.01 betas=0.9,0.999,0.9999 eps=0.000001", | |
| "output_dir": "/data/Train/output", | |
| "output_name": "kohya_flux_miltiple_resolution", | |
| "persistent_data_loader_workers": true, | |
| "pretrained_model_name_or_path": "/data/models/flux1-dev.safetensors", | |
| "prior_loss_weight": 1, | |
| "random_crop": false, | |
| "rank_dropout": 0, | |
| "rank_dropout_scale": false, | |
| "reg_data_dir": "", | |
| "rescaled": false, | |
| "resume": "", | |
| "resume_from_huggingface": "", | |
| "sample_every_n_epochs": 0, | |
| "sample_every_n_steps": 0, | |
| "sample_prompts": "", | |
| "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": "fp16", | |
| "save_state": false, | |
| "save_state_on_train_end": true, | |
| "save_state_to_huggingface": false, | |
| "scale_v_pred_loss_like_noise_pred": false, | |
| "scale_weight_norms": 0, | |
| "sdxl": false, | |
| "sdxl_cache_text_encoder_outputs": false, | |
| "sdxl_no_half_vae": false, | |
| "seed": 17415, | |
| "shuffle_caption": false, | |
| "split_mode": false, | |
| "split_qkv": false, | |
| "stop_text_encoder_training_pct": 0, | |
| "t5xxl": "/data/models/clip/t5xxl_fp8_e4m3fn.safetensors", | |
| "t5xxl_max_token_length": 512, | |
| "text_encoder_lr": 5e-05, | |
| "timestep_sampling": "shift", | |
| "train_batch_size": 1, | |
| "train_blocks": "single", | |
| "train_data_dir": "", | |
| "train_norm": false, | |
| "train_on_input": true, | |
| "training_comment": "", | |
| "unet_lr": 1e-04, | |
| "unit": 1, | |
| "up_lr_weight": "", | |
| "use_cp": false, | |
| "use_scalar": false, | |
| "use_tucker": false, | |
| "v2": false, | |
| "v_parameterization": false, | |
| "v_pred_like_loss": 0, | |
| "vae": "", | |
| "vae_batch_size": 0, | |
| "wandb_api_key": "", | |
| "wandb_run_name": "", | |
| "weighted_captions": false, | |
| "xformers": "sdpa" | |
| } |
