WhiteAlbum2 Lora Wan2.2
详情
下载文件
模型描述
触发词:WhiteAlbum2
这是一个专门训练的White Album 2风格LoRA,可还原该系列中的经典角色。当结合角色特征描述与对应角色名称(如Setsuna、Kazusa、Haruki)作为触发词时,可生成对应角色的外观。
使用示例:
anime style, WhiteAlbum2, kazusa, A young woman with long straight black hair and striking blue eyes stands at a modern minimalist kitchen counter, wearing a crisp white button-up shirt with long sleeves, a wide red belt with a gold buckle, and a dark houndstooth-patterned skirt; sunlight pours through the window onto white cabinets, a marble countertop, and green plants; she carefully tilts a silver coffee pot, letting a thin stream of hot coffee flow slowly into the cup she steadies with her other hand, her expression focused and calm; the camera captures a medium close-up with a lens push towards her hands, the sunny daylight creates warm soft side lighting that highlights the steam rising; the aesthetic is naturalistic and serene, stylized in a cinematic slice-of-life realism.
赞助我一杯咖啡 https://ko-fi.com/dreamweebs
工作流测试:
我测试了两种主要工作流:
增强版工作流:集成HPS奖励LoRA与Light2V I2V 64 WAN2.1(未测试WAN2.2)
- 优点:HPS显著提升细节质量
- 缺点:Light2V易导致前几帧场景重叠;HPS引入一些异常色彩伪影
原生工作流:未使用任何加速LoRA
- 优点:输出相对稳定
- 缺点:生成速度较慢
使用建议:
- 提示词至关重要:简单提示词常导致静态画面,强烈建议使用AI辅助撰写详细提示词
- 工作流选择:不同工作流设置及是否添加Light2V加速LoRA会显著影响最终效果
技术说明:本LoRA基于WAN2.2训练。初期因Light2V权重设置错误(设为3),导致所有输出出现雪花噪点。经近期调整后,在新工作流下表现良好。虽可能存在因训练轮次不足导致的轻微问题,但已完全满足一般使用需求。
