{
    "model_id": 692245,
    "model_name": "XL-Vibratory roller-hans_miao",
    "model_description": "<p>\n 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。\n <br/>\n This is an XL model trained on 5 million images, featuring over 2,000 strong style tags.\n</p>\n<p>\n 模型的训练过程包括了多阶段训练：\n <br/>\n The training process includes multiple stages:\n</p>\n<ul>\n <li>\n  <p>\n   使用A40双卡训练了400小时\n  </p>\n </li>\n <li>\n  <p>\n   Trained for 400 hours with dual A40 GPUs\n  </p>\n </li>\n <li>\n  <p>\n   通过4090训练了1104小时来完善底模\n  </p>\n </li>\n <li>\n  <p>\n   Enhanced with 1104 hours of training using the 4090 GPU\n  </p>\n </li>\n <li>\n  <p>\n   最后在双卡A100 80G上进行了504小时的画风训练\n  </p>\n </li>\n <li>\n  <p>\n   Finally, completed 504 hours of style training on dual A100 80G GPUs\n  </p>\n </li>\n</ul>\n<p>\n <strong>\n  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。\n </strong>\n <br/>\n <strong>\n  Version 1.0 still has some underfitted tags, which will be fixed in future updates.\n </strong>\n</p>\n<p>\n 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 和复读机bot 在训练集构建方面的帮助。\n <br/>\n Special thanks to rnglg2 for computational and data processing support, and to 且听风吟, Willy, 青秋, nano, Cyanelis Deuxieme, and 复读机bot for assistance in building the training dataset.\n</p>\n<hr/>\n<h3 id=\"(usage-recommendation)-k33ss9hsh\">\n 模型的使用建议 (Usage Recommendation)\n</h3>\n<p>\n 推荐CFG不超过5\n</p>\n<p>\n 我们对数据集进行了美学评分，评分标准如下：\n <br/>\n We applied aesthetic scoring to the dataset, with the following rating criteria:\n</p>\n<ul>\n <li>\n  <p>\n   Core &gt; 0.75: 质量标签 = \"masterpiece\"\n  </p>\n </li>\n <li>\n  <p>\n   Core &gt; 0.75: Quality Tag = \"masterpiece\"\n  </p>\n </li>\n <li>\n  <p>\n   0.6 &lt; score &lt;= 0.75: 质量标签 = \"high quality\"\n  </p>\n </li>\n <li>\n  <p>\n   0.6 &lt; score &lt;= 0.75: Quality Tag = \"high quality\"\n  </p>\n </li>\n <li>\n  <p>\n   0.5 &lt; score &lt;= 0.6: 质量标签 = \"normal quality\"\n  </p>\n </li>\n <li>\n  <p>\n   0.5 &lt; score &lt;= 0.6: Quality Tag = \"normal quality\"\n  </p>\n </li>\n <li>\n  <p>\n   0.3 &lt; score &lt;= 0.5: 质量标签 = \"low quality\"\n  </p>\n </li>\n <li>\n  <p>\n   0.3 &lt; score &lt;= 0.5: Quality Tag = \"low quality\"\n  </p>\n </li>\n <li>\n  <p>\n   score &lt;= 0.3: 质量标签 = \"worst quality\"\n  </p>\n </li>\n <li>\n  <p>\n   Score &lt;= 0.3: Quality Tag = \"worst quality\"\n  </p>\n </li>\n</ul>\n<p>\n 正常使用时，只需在标签前添加\n <code>\n  masterpiece\n </code>\n ,\n <code>\n  best quality\n </code>\n , 或\n <code>\n  high quality\n </code>\n 即可。\n <br/>\n For normal use, simply add\n <code>\n  masterpiece\n </code>\n ,\n <code>\n  best quality\n </code>\n , or\n <code>\n  high quality\n </code>\n before the tag.\n</p>\n<p>\n 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。\n <br/>\n The full table of style tags will be provided in the full release. For now, you can refer to the example images.\n</p>\n<p>\n <a href=\"https://rnglg2-my.sharepoint.com/:u:/g/personal/hans_rnglg2_onmicrosoft_com/EfcBzxcz06xAmOpPrQHjvb8B2e3dUpwcbckEU1UC5Rm0fw?e=f7aIiC\" rel=\"ugc\" target=\"_blank\">\n  https://rnglg2-my.sharepoint.com/:u:/g/personal/hans_rnglg2_onmicrosoft_com/EfcBzxcz06xAmOpPrQHjvb8B2e3dUpwcbckEU1UC5Rm0fw?e=f7aIiC\n </a>\n</p>\n<p>\n 训练参数 (Training Parameters)\n</p>\n<p>\n resolution = \"1024,1024\"\n</p>\n<p>\n enable_bucket = true\n</p>\n<p>\n min_bucket_reso = 256\n</p>\n<p>\n max_bucket_reso = 1536\n</p>\n<p>\n bucket_reso_steps = 32\n</p>\n<p>\n output_dir = \"/root/\"\n</p>\n<p>\n save_model_as = \"safetensors\"\n</p>\n<p>\n save_precision = \"fp16\"\n</p>\n<p>\n save_every_n_epochs = 2\n</p>\n<p>\n max_train_epochs = 20\n</p>\n<p>\n train_batch_size = 5\n</p>\n<p>\n gradient_checkpointing = false\n</p>\n<p>\n learning_rate = 0.00003\n</p>\n<p>\n learning_rate_te1 = 0.000001\n</p>\n<p>\n learning_rate_te2 = 0.000001\n</p>\n<p>\n lr_scheduler = \"cosine_with_restarts\"\n</p>\n<p>\n lr_scheduler_num_cycles = 20\n</p>\n<p>\n optimizer_type = \"AdamW\"\n</p>\n<p>\n min_snr_gamma = 5\n</p>\n<p>\n sample_every_n_epochs = 1\n</p>\n<p>\n log_with = \"tensorboard\"\n</p>\n<p>\n logging_dir = \"./logs\"\n</p>\n<p>\n caption_extension = \".txt\"\n</p>\n<p>\n shuffle_caption = true\n</p>\n<p>\n weighted_captions = false\n</p>\n<p>\n keep_tokens = 4\n</p>\n<p>\n max_token_length = 255\n</p>\n<p>\n multires_noise_iterations = 8\n</p>\n<p>\n multires_noise_discount = 0.4\n</p>\n<p>\n no_token_padding = false\n</p>\n<p>\n mixed_precision = \"bf16\"\n</p>\n<p>\n full_bf16 = true\n</p>\n<p>\n xformers = true\n</p>\n<p>\n lowram = false\n</p>\n<p>\n cache_latents = true\n</p>\n<p>\n cache_latents_to_disk = true\n</p>\n<p>\n persistent_data_loader_workers = true\n</p>\n<p>\n train_text_encoder = true\n</p>\n<h3 id=\"(disclaimer)-qfv7rfgcx\">\n 免责声明 (Disclaimer)\n</h3>\n<p>\n 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。\n <br/>\n 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.\n</p>\n<h3 id=\"(license)-rigvdupl1\">\n 许可证 (License)\n</h3>\n<p>\n 许可证：Fair AI Public License 1.0-SD\n</p>\n<p>\n</p>\n<p>\n</p>\n",
    "model_url": "https://civitai.com/models/692245",
    "model_type": "Checkpoint",
    "model_tags": [
        "anime",
        "base model",
        "woman",
        "game character",
        "girls",
        "video game"
    ],
    "download_link": "https://civitai.com/api/download/models/774719",
    "preview_image_url": "https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/198a6ee0-c10f-49bc-a614-6400e22c1c3f/width=450/26402475.jpeg",
    "model_version_id": null,
    "model_version_download_link": null,
    "model_version_image_url": null
}