Yat-Sen-Azur-Lane-Paragon-of-Celestial-Grace

我要开发同款
匿名用户2024年07月31日
16阅读
开发技术Pytorch
所属分类ai、text-to-image、lora
开源地址https://modelscope.cn/models/sd_lora/Yat-Sen-Azur-Lane-Paragon-of-Celestial-Grace
授权协议Apache License 2.0

作品详情

Yet-San's new costume in 2023 spring.

She is wearing a kind of traditional Chinese dress called ru skirt(襦ru裙qun).

===================

V3:

You can take off her clothes even at strength 1√

即使权重1,衣服可脱√

If you want to change her clothes easily, recommend below 0.85.

想要比较容易地换衣服,建议权重低于0.85

Multired noise and network dropout enabled.

金字塔噪声和dropout也开了。

I used AI-generated imgs to train V2 by mistake which cause the style to be overbaked. As for V3 I removed AI-generated imgs and add more imgs to reduce style overbaking.

V2错误地使用了AI图作为素材导致画风过拟,V3移除了AI图并加入更多素材以减少画风过拟。

===================

V2:

More training data and steps, deleted some tags in training,

更多训练数据和步数,删掉了一些训练的tag

Now there's a trigger word.

现在有触发词了。

The trigger word is not necessary If you want to change her clothes more easily.

如果你想更轻松地换衣服,触发词不是必要的。

A little bit overfitting which might influence your models' style with weight 1, so you'd better use it with weight around 0.8.

有一丢丢过拟,在权重为1时会影响画风,所以你最好在0.8左右使用。

Suggested Prompt(推荐提示词): black hair, white legwear, white_skirt, red eyes……

Suggested Prompts are not necessary, but you can get more stable outputs especially when the weight is low or you are using realistic models which this anime lora wasn't made for.

推荐提示词不是必要的,但可以让你得到更稳定的结果,特别是权重较低或你要用在不是为此训练的真人模型上时。

===================

V1 Suggested Prompt: black hair, white dress, white legwear, cleavage

声明:本文仅代表作者观点,不代表本站立场。如果侵犯到您的合法权益,请联系我们删除侵权资源!如果遇到资源链接失效,请您通过评论或工单的方式通知管理员。未经允许,不得转载,本站所有资源文章禁止商业使用运营!
下载安装【程序员客栈】APP
实时对接需求、及时收发消息、丰富的开放项目需求、随时随地查看项目状态

评论