🚀 头顶摇臂LoRA模型(适用于Wan2.1 14B I2V 480p)
本模型可实现平滑地从拍摄对象上方升起,以俯视角展示场景,非常适合用于过渡、戏剧性展示或优雅地转换视角。
🚀 快速开始
本模型基于Wan2.1 14B I2V 480p模型进行训练,你可以按照以下步骤开始使用:
- 下载模型文件和推理工作流文件。
- 按照推荐设置进行参数配置。
- 使用触发词和提示模板生成视频。
✨ 主要特性
- 基于基础模型训练:基于Wan2.1 14B 480p I2V基础模型进行训练。
- 结果一致性高:在不同对象类型上都能产生一致的结果。
- 提示结构简单:提示结构简单,易于调整。
📦 安装指南
📥 下载链接
💻 使用示例
示例提示
- 示例1
- 提示文本:A young man with curly hair and wire frame wings stands in an abandoned building, looking directly at the camera. The camera begins a cr4n3 crane over the head movement, rising smoothly upwards. As the cr4n3 crane over the head movement progresses, the young man looks up towards the camera. The cr4n3 crane over the head movement concludes with a high - angle shot, looking down at the young man with his wire frame wings.
- 输出视频:示例视频1
- 示例2
- 提示文本:A man with a white beard, wearing a red bandana, a black leather jacket, an orange t - shirt with "REMADE AI" printed on it, and blue jeans, sits on a black motorcycle, looking directly at the camera. The camera then performs a cr4n3 crane over the head movement, smoothly rising higher and tilting downwards, revealing the man and motorcycle against a dimly lit garage backdrop. As the cr4n3 crane over the head continues, the man looks up towards the rising camera.
- 输出视频:示例视频2
- 示例3
- 提示文本:A person with short, textured hair, wearing a graphic t - shirt, a choker necklace, and baggy jeans, crouches on a wet, reflective surface. The background features vertical neon lights in red and blue. The camera begins a cr4n3 crane over the head movement, smoothly rising upwards and tilting down, offering an increasingly high - angle view of the person and the rain - slicked ground. As the cr4n3 crane over the head movement progresses, the person looks up towards the rising camera.
- 输出视频:示例视频3
推荐设置
属性 |
详情 |
LoRA强度 |
1.0 |
嵌入式引导比例 |
6.0 |
流偏移 |
5.0 |
触发词
关键触发短语为:cr4n3 crane over the head movement
提示模板
可参考上述示例提示,这种提示方式效果较好。
ComfyUI工作流
本LoRA模型可与 Kijai的Wan Video Wrapper工作流 的修改版本配合使用。主要修改是添加了一个连接到基础模型的Wan LoRA节点。
修改后的工作流可在上述下载部分找到。
📚 详细文档
模型信息
模型权重以Safetensors格式提供,可在上述下载部分找到。
训练详情
属性 |
详情 |
基础模型 |
Wan2.1 14B I2V 480p |
训练数据 |
使用50秒的视频进行训练,该视频由10个短剪辑组成(每个剪辑单独添加字幕),这些场景都使用了头顶摇臂相机运动。 |
训练轮数 |
50 |
额外信息
训练使用了 Diffusion Pipe for Training。
致谢
特别感谢Kijai提供的ComfyUI Wan Video Wrapper和tdrussell提供的训练脚本!
🔗 社区
📄 许可证
本项目采用Apache - 2.0许可证。