1 基本介绍（Introduction）

基于 kohaku-epsilon-rev2 训练，权重调整为 1 即可，使用了 1172 张自己在 NovelAI V3 中生成的图像，训练参数如下：

Trained based on kohaku-epsilon-rev2, setting the weight to 1 is fine. Used 1172 images generated by myself in NovelAI V3. The training parameters are as follows:

model_train_type = "sdxl-lora"
pretrained_model_name_or_path = "E:/AI/lora-scripts-v1.8.5/sd-models/kohaku-xl-epsilon-rev2.safetensors"
v2 = false
train_data_dir = "E:/AI/lora-scripts-v1.8.5/train/5_style5"
prior_loss_weight = 1
resolution = "1024,1024"
enable_bucket = true
min_bucket_reso = 512
max_bucket_reso = 1536
bucket_reso_steps = 64
output_name = "hxtest1"
output_dir = "./output"
save_model_as = "safetensors"
save_precision = "bf16"
save_every_n_epochs = 1
max_train_epochs = 10
train_batch_size = 1
gradient_checkpointing = false
network_train_unet_only = false
network_train_text_encoder_only = false
learning_rate = 0.0002
unet_lr = 0.00001
text_encoder_lr = 0.00001
lr_scheduler = "cosine"
lr_warmup_steps = 3000
optimizer_type = "PagedAdamW8bit"
network_module = "lycoris.kohya"
network_dim = 64
network_alpha = 32
train_norm = false
log_with = "tensorboard"
logging_dir = "./logs"
caption_extension = ".txt"
shuffle_caption = true
keep_tokens = 0
max_token_length = 255
seed = 1337
mixed_precision = "bf16"
xformers = true
lowram = false
cache_latents = true
cache_latents_to_disk = true
persistent_data_loader_workers = true
network_args = [ "conv_dim=4", "conv_alpha=1", "dropout=0", "algo=locon" ]

经过测试，感觉最终的版本综合效果更好。

2 模型测试（model test）

2-1 概述（overview）：

模型对各个底模的效果测试结果如下：

kohakuXL-epsilon-rev2：最好使用 kohakuXL-epsilon-rev2（如果有更新版本的 kohakuXL 出了也可以试试）；
animagineXL：在 animagineXLV31 中测试效果也很明显，本人推测该模型也可以用于以 animagineXL 为底模的各个模型中；
ponyDiffusionV6XL：经过多次测试，该模型在 ponyDiffusionV6XL 中效果并不明显，该风格已经有人做出来了，名字为 tPonynai3，效果很好；
tPonynai3_v4：我也以 tPonynai3 作为底模进行了测试，感觉该模型对其生成结果会产生一定的影响（有好有坏），如果使用 tPonynai3 作为底模生成图片，可以尝试一下本模型~
本模型对 anythingXL 也在一定程度上适用，不过该模型本身就很可爱，所以效果可能不太明显。

加入本模型可能会导致生成的内容出现变化。

The testing results of the model on various base models are as follows:

kohakuXL-epsilon-rev2: It is best to use kohakuXL-epsilon-rev2 (if a newer version of kohakuXL is available, it is worth trying).
animagineXL: The test results in animagineXLV31 were also significant. I speculate that this model can also be used in various models based on animagineXL.
ponyDiffusionV6XL: After multiple tests, the effect of this model in ponyDiffusionV6XL is not obvious. This style has already been achieved by someone else, named tPonynai3, with excellent results.
tPonynai3_v4: I also tested with tPonynai3 as the base model and found that this model can influence its generation results (both positively and negatively). If using tPonynai3 as the base model to generate images, it is worth trying this model.
anythingXL: This model is also somewhat applicable to anythingXL, but since this model is inherently very cute, the effect may not be very noticeable.