这是一个经过500万张图像训练而成的XL大模型，内置了超过2000个强风格标签。
This is an XL model trained on 5 million images, featuring over 2,000 strong style tags.
模型的训练过程包括了多阶段训练：
The training process includes multiple stages:
使用A40双卡训练了400小时
Trained for 400 hours with dual A40 GPUs
通过4090训练了1104小时来完善底模
Enhanced with 1104 hours of training using the 4090 GPU
最后在双卡A100 80G上进行了800小时的画风训练
Finally, completed 504 hours of style training on dual A100 80G GPUs
目前正式版1.0仍有部分标签欠拟合，后续版本会进行修复。
Version 1.0 still has some underfitted tags, which will be fixed in future updates.
特别感谢rnglg2在算力和数据处理方面的支持，以及且听风吟、Willy、青秋、nano、Cyanelis Deuxieme、群阿没、复读机bot在训练集构建方面的帮助。
Special thanks to rnglg2 for computational and data processing support, and to 且听风吟, Willy, 青秋, nano, 群阿没、Cyanelis Deuxieme, and 复读机bot for assistance in building the training dataset.
训练底模来自【SDXL】/Anime/bulldozer_BETA - v2.0 | Stable Diffusion XL Checkpoint | Civitai
模型的使用建议 (Usage Recommendation)
推荐CFG不超过5
我们对数据集进行了美学评分，评分标准如下：
We applied aesthetic scoring to the dataset, with the following rating criteria:
Core > 0.75: 质量标签 = "masterpiece"
Core > 0.75: Quality Tag = "masterpiece"
0.6 < score <= 0.75: 质量标签 = "high quality"
0.6 < score <= 0.75: Quality Tag = "high quality"
0.5 < score <= 0.6: 质量标签 = "normal quality"
0.5 < score <= 0.6: Quality Tag = "normal quality"
0.3 < score <= 0.5: 质量标签 = "low quality"
0.3 < score <= 0.5: Quality Tag = "low quality"
score <= 0.3: 质量标签 = "worst quality"
Score <= 0.3: Quality Tag = "worst quality"
正常使用时，只需在标签前添加masterpiece, best quality, 或high quality即可。
For normal use, simply add masterpiece, best quality, or high quality before the tag.
风格标签的完整表格将在完整版发布时提供，目前可以通过参考示例图来使用。
The full table of style tags will be provided in the full release. For now, you can refer to the example images.
训练参数 (Training Parameters)
resolution = "1024,1024"
enable_bucket = true
min_bucket_reso = 256
max_bucket_reso = 1536
bucket_reso_steps = 32
output_dir = "/root/"
save_model_as = "safetensors"
save_precision = "fp16"
save_every_n_epochs = 2
max_train_epochs = 20
train_batch_size = 5
gradient_checkpointing = false
learning_rate = 0.00003
learning_rate_te1 = 0.000001
learning_rate_te2 = 0.000001
lr_scheduler = "cosine_with_restarts"
lr_scheduler_num_cycles = 20
optimizer_type = "AdamW"
min_snr_gamma = 5
sample_every_n_epochs = 1
log_with = "tensorboard"
logging_dir = "./logs"
caption_extension = ".txt"
shuffle_caption = true
weighted_captions = false
keep_tokens = 4
max_token_length = 255
multires_noise_iterations = 8
multires_noise_discount = 0.4
no_token_padding = false
mixed_precision = "bf16"
full_bf16 = true
xformers = true
lowram = false
cache_latents = true
cache_latents_to_disk = true
persistent_data_loader_workers = true
train_text_encoder = true
免责声明 (Disclaimer)
请遵守当地法律法规，以免造成麻烦。鉴于模型的实际用途不受模型作者控制，因模型输出的图片所产生的一切后果由图片输出者自行承担。基于此模型训练出的衍生模型请标明出处。
As the actual use of the model is beyond the control of the model creators, all consequences arising from images generated by this model are the sole responsibility of the user.