🚀 用于Flux的凶狠高大龙LoRA模型
本项目是一个基于LoRA技术的模型,专为Flux.1-dev模型微调,能够生成更加凶狠、高大的龙形象,解决了原模型在表现龙的规模和姿态方面的难题。
🚀 快速开始
激活方式
无需触发词,可使用以下提示词:
- 一只巨大凶狠的红龙高耸在一个小骑士之上
- 一座山峰上的巨型黑龙
推荐训练轮数
使用12 - 15轮训练的文件(已提供),建议从14或15轮开始。
LoRA强度
Clip和模型强度设置在0.8到1.0之间。
分辨率
建议使用垂直分辨率(例如768x1344)以展现规模感。
✨ 主要特性
本LoRA模型针对Flux.1-dev进行微调,能够生成更凶狠、更具威胁性且体型更大的龙,使其高耸于人物或风景之上,有效解决了Flux.1-dev在表现规模和姿态方面的挑战。
📦 安装指南
文档未提及安装步骤,暂不提供。
💻 使用示例
基础用法
使用如下提示词即可生成相应的龙图像:
massive menacing red dragon towering over a tiny knight
高级用法
可根据不同的场景和需求,调整提示词的细节,如龙的颜色、姿态,人物的穿着等,以生成更符合预期的图像。例如:
A photo-realistic shoot from a profile camera angle about a young girl with long, flowing blonde hair standing in front of a large, menacing dragon. the girl, who appears to be in her early twenties, is facing away from the camera, with her back to the viewer. she is wearing a red dress with ruffled sleeves, and her hair is styled in a braid. the dragon, with its sharp teeth and orange scales, has a menacing expression, and its eyes are focused intently on the girl. the background is blurred, with rocks and greenery visible, and the lighting is dramatic, casting shadows on the dragon's face and body. the overall mood is dark and moody, with a focus on the character and the dragon.
📚 详细文档
模型描述
本LoRA模型针对Flux.1-dev进行微调,旨在生成更凶狠、更具威胁性且体型更大的龙,使其高耸于人物或风景之上,解决了Flux.1-dev在表现规模和姿态方面的挑战。
标题风格
示例标题风格展示了如何通过详细的描述,营造出特定的场景氛围,如:
A photo-realistic shoot from a profile camera angle about a young girl with long, flowing blonde hair standing in front of a large, menacing dragon. the girl, who appears to be in her early twenties, is facing away from the camera, with her back to the viewer. she is wearing a red dress with ruffled sleeves, and her hair is styled in a braid. the dragon, with its sharp teeth and orange scales, has a menacing expression, and its eyes are focused intently on the girl. the background is blurred, with rocks and greenery visible, and the lighting is dramatic, casting shadows on the dragon's face and body. the overall mood is dark and moody, with a focus on the character and the dragon.
关键训练参数
参数 |
详情 |
引擎 |
kohya |
训练轮数 |
17 |
重复次数 |
6 |
步数 |
1887 |
分辨率 |
1024 |
Unet学习率 |
0.00020 |
学习率调度器 |
cosine_with_restarts |
优化器 |
AdamW8bit |
网络维度(秩) |
64 |
网络Alpha |
32 |
🔧 技术细节
文档未提供具体的技术实现细节,暂不展示。
📄 许可证
本模型以MIT许可证发布,可自由使用、修改和分发,但请注明来源。
模型下载
本模型的权重以Safetensors格式提供。
点击下载(在文件与版本标签中)。