🚀 Wan-LoRA-Arcane-Jinx-v1

本项目是一个文本到图像/视频的LoRA模型，基于Wan-AI/Wan2.1-T2V-1.3B基础模型训练，可根据输入的文本描述生成特定风格的图像或视频。

🚀 快速开始

模型信息

示例展示

• 输入文本：'csetiarcane A medium shot frames a J1nx in neon-washed leather, her silhouette sharp against flickering signs. Her long blue hair catches blue light as she moves sideways through the frame. Her mesh top gleams beneath a studded jacket. Behind her, steam rises from street vents into the rainy night.'
  • 输出视频：点击查看
• 输入文本：'csetiarcane Nfj1nx stands at the edge of a candlelit rooftop terrace overlooking glittering city towers. The camera faces her from the front, capturing every detail of her gentle smile as she meets the gaze of her partner—dressed in a sleek black blazer, collar slightly undone, and he has sporting jet-black hair that frames his face. Soft string lights sway overhead, highlighting her long blue hair and the warm reflection in her eyes. She takes a small step forward, and he carefully places one hand on her cheek before leaning in for a tender kiss, their silhouettes merging at the center of the frame.'
  • 输出视频：点击查看
• 输入文本：'csetiarcane Nfj1nx moves along a tree-shaded park path, coming directly toward the camera. A matte-black pair of over-ear headphones partially covers her ears, letting most of her bright blue hair remain visible. She wears a simple white athletic hoodie paired with black biker shorts. Pastel-toned running shoes quietly crunch on the gravel, reflecting her relaxed yet confident rhythm.'
  • 输出视频：点击查看
• 输入文本：'csetiarcane Nfj1nx moves along a tree-shaded park path, coming directly toward the camera. A matte-black pair of over-ear headphones partially covers her ears, letting most of her bright blue hair remain visible. She wears a simple white athletic hoodie paired with black biker shorts. Pastel-toned running shoes quietly crunch on the gravel, reflecting her relaxed yet confident rhythm.'
  • 输出视频：点击查看

✨ 主要特性

• 本模型采用先在图像上训练LoRA，再在视频上继续训练的方法。
• 模型偏好长描述性提示词，使用短提示词时LoRA效果较弱。

🔧 技术细节

训练参数

• 学习率（LR）：2e-5
• 优化器：adamw_optimi
• 训练轮数（epochs）：图像训练15轮 + 视频训练7轮
• 训练步数（steps）：图像训练2550步 + 视频训练2520步
• 数据集：33张图像和35个视频
• 重复次数：10
• 秩（rank）：128
• 批量大小（batch size）：2
• 梯度累积步数：1

训练代码仓库

训练使用了 Diffusion-Pipe 仓库。

💻 使用示例

推理工具推荐

• ComfyUI原生节点
• Kijai的ComfyUI包装节点

提示词建议

• 触发词：csetiarcane, Nfj1nx, 蓝色头发
• 强度：0.9 - 1.0
• 如果仅使用触发词无法得到所需风格或角色，建议在提示词中添加 "animation style"，但可能会使结果过于卡通化。

📄 许可证

本LoRA是作为粉丝项目的一部分创建的，仅用于研究目的，不用于商业用途。它基于受版权保护的电视剧《Arcane》。用户自行承担使用该模型的风险，有义务遵守版权法和适用法规。该模型是为非商业目的开发的，无意侵犯任何版权。对于因使用该模型而产生的任何损害或法律后果，本人不承担任何责任。

👏 致谢

特别感谢 Kijai 为社区所做的不懈努力和出色工作，使我们能够尽快使用这些解决方案；也感谢 Comfyanonymous 如此迅速且高质量地将其原生集成到ComfyUI中；还要感谢 TDRussel 如此迅速地让社区能够进行LoRA训练。