斗破苍穹_萧薰儿_中洲/Battle_Through The Heavens/xiaoxuner_zhongzhou

推荐使用2.0，2.0单一lora比1.0 分开outfit + face lora一致性效果更好

Version 2.0 is recommended; the single LoRA in 2.0 delivers better consistency than the separate 1.0 LoRAs.

1. 触发词 / Trigger Words

<lora:xiaoxuner_zhongzhou:0.8>, xiaoxuner_zhongzhou

推荐搭配提示词：

Recommended prompt tags:

black hair, dress, detached sleeves, ponytail, hair ornament, jewelry, earrings, long hair

2. VAE / VAE

使用 SDXL VAE，可在 Hugging Face 搜索下载 sdxlVAE_sdxlVAE.safetensors。

Recommended VAE: search Hugging Face for sdxlVAE_sdxlVAE.safetensors.

3. 数据集说明 / Dataset

训练集调整为 300+ mixed dataset，混合 face、outfit、head ornament、hair、full-body silhouette 等多种构图。数据集质量决定 90% 的 LoRA 效果。

The training dataset contains 300+ mixed images, covering face, outfit, head ornament, hair, and full-body silhouette. Dataset quality determines 90% of the final LoRA quality.

4. 训练参数 / Training Config Toml

model_train_type = "sdxl-lora"
pretrained_model_name_or_path = "path to your base model"
resume = ""
vae = "path to your VAE_sdxlVAE.safetensors"
train_data_dir = "path to your dataset"
prior_loss_weight = 1
resolution = "1024,1024"
enable_bucket = true
min_bucket_reso = 512
max_bucket_reso = 2048
bucket_reso_steps = 64
bucket_no_upscale = false
output_name = "xiaoxuner_zhongzhou_illustrious"
output_dir = "path to your lora output path"
save_model_as = "safetensors"
save_precision = "bf16"
save_every_n_epochs = 2
save_state = true
save_last_n_epochs_state = 2
max_train_epochs = 14
train_batch_size = 2
gradient_checkpointing = true
gradient_accumulation_steps = 1
network_train_unet_only = true
network_train_text_encoder_only = false
learning_rate = 0
unet_lr = 0.00007
text_encoder_lr = 0.000003
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 100
optimizer_type = "AdamW8bit"
optimizer_args = [ "weight_decay=0.01" ]
network_module = "networks.lora"
network_dim = 64
network_alpha = 64
randomly_choice_prompt = false
positive_prompts = "xiaoxuner_zhongzhou, solo, black hair, dress, detached sleeves, ponytail, hair ornament, jewelry, earrings, long hair"
negative_prompts = "lowres, bad anatomy, bad hands, bad face, bad eyes, deformed, distorted, text, logo, watermark, signature, cropped, blurry, jpeg artifacts, extra fingers, missing fingers, monochrome, grayscale"
sample_width = 896
sample_height = 1600
sample_cfg = 6
sample_seed = 2333
sample_steps = 30
sample_sampler = "dpmsolver++"
sample_every_n_epochs = 2
log_with = "tensorboard"
logging_dir = "./logs"
caption_extension = ".txt"
shuffle_caption = true
keep_tokens = 1
keep_tokens_separator = ""
max_token_length = 255
color_aug = false
flip_aug = false
random_crop = false
seed = 1337
mixed_precision = "bf16"
no_half_vae = true
xformers = true
lowram = false
cache_latents = true
cache_latents_to_disk = false
persistent_data_loader_workers = true
vae_batch_size = 1

5. ADetailer 脸部增强
推荐启用 ADetailer Face Detection，并使用相同 LoRA 进行脸部细节修复。推荐参数：DPM++ 2M SDE Karras、1024×1024、Denoise 0.65，ADetailer Prompt 示例：

It is recommended to enable ADetailer Face Detection and use the same LoRA for face enhancement. Recommended settings: DPM++ 2M SDE Karras, 1024×1024, Denoise 0.65. Example ADetailer Prompt:

<lora:xiaoxuner_zhongzhou:1>, xiaoxuner_zhongzhou

6. Principle
LoRA 的本质是在原模型权重上学习低秩增量，同一个角色，脸部、发型、头饰、服装和整体轮廓在 latent space 中不是完全独立的概念，而是同一个角色分布的耦合特征。拆成 face 和 outfit，本质上是在同一次生成中叠加两个不同的权重增量，学到的角色分布不完全一致时，容易产生 identity drift、feature conflict 或 ADetailer 局部重绘后脸部不一致，统一lora让模型学习一个共同的角色增量，因此主图生成和adetailer修脸都沿同一个角色分布优化，测试下来获得更高的人物还原度和整体一致性。

The essence of LoRA is to learn low-rank updates on top of the base model weights. For a given character, facial features, hairstyle, accessories, outfit, and overall silhouette are not completely independent concepts in latent space, but rather coupled features belonging to the same character distribution. Training separate face and outfit LoRAs effectively learns two different weight updates. When the learned character distributions do not fully align, this can lead to identity drift, feature conflicts, or inconsistencies after ADetailer face refinement. A unified Character LoRA allows the model to learn a shared character update, enabling both the main image generation and ADetailer face enhancement to optimize within the same character distribution. In practice, this approach consistently produces higher character fidelity and stronger overall visual consistency.