breast beaten
详情
下载文件
关于此版本
模型描述
v2.0
提升了通用性，降低了过拟合问题，不再必须通过Hires或图生图方式间接出图，但使用Hires仍能显著提升图像质量。
推荐权重：0.8。0.6~1.0均有不错效果，在使用Pony XL v6原模型时可使用1.0权重，但加入其他风格LoRA后可能需要降低该LoRA或风格LoRA的权重，例如Concept Art Eclipse Style LoRA_Pony XL v6就需要降低权重，且效果很差；而Expressive_H则可同时以1.0权重使用。其他Pony模型尚未测试。
触发词：breastbeaten
推荐提示词：1girl, 1boy, large breasts, breastbeaten
可选相关提示词：nipple slip, areola slip, pov hand
添加nipple slip时，乳头会出现在拳头边缘，areola slip同理。
若不添加pov hand，有时会出现愤怒的男性面部，可根据需求加入pov hand等提示进行控制。
提高了泛用性，降低了过拟合的问题，通过Hires或者图生图的方式曲线出图不再必须了，但使用Hires还是能很大程度提高图像质量。
推荐权重：0.8。0.6~1.0都有一定效果，在使用pony xl v6原模型的情况下可以使用到1.0权重，但加入其他风格lora后可能需要下调该lora或风格lora权重，如Concept Art Eclipse Style LoRA_Pony XL v6就需要下调权重，并且效果很差，但如Expressive_H可以两者都使用在1.0权重。其余pony模型没有测试。
触发词：breastbeaten
推荐提示词：1girl, 1boy, large breasts, breastbeaten
可选相关提示词：nipple slip, areola slip, pov hand
加上nipple slip表现为nipple在拳头边缘，areola slip同理
如果不加pov hand有时会出现angry male face，可以根据需求加入pov hand等控制
Improved applicability, reduced overfitting problems, Hires and img2img are no longer necessary, but Hires is still a significant improvement in image quality.
Recommended weights: 0.8,0.6~1.0. A weight of 1.0 can be used with the pony xl v6 original model, but when adding other style Lora, the Lora or style Lora weight may need to be reduced. For instance, Concept Art Eclipse Style LoRA_Pony XL v6 requires a reduction in weight, and the performerce is awful, whereas Expressive_H can use both at 1.0 weight. Other pony models have not been tested.
Trigger word: breastbeaten
Recommended prompt: 1girl, 1boy, large breasts, breastbeaten
nipple slip, areola slip, pov hand
If you add nipple slip, the nipple is at the edge of the fist, and the same is true of areola slip
If you do not add pov hand, sometimes there will be an angry male face. you can add pov hand or other prompts to control it.
=================================
v1.0
第一次训练lora，调整了3次终于有点作用了
效果和适用性很差，应该跟训练集和训练有点过拟合相关。不过按示例图提示词结合高清修复还是能抽几张卡的
The first time training LoRa, after adjusting it three times, it finally started to have some effect. The model's performance and applicability are quite poor, likely related to overfitting and training sets. However, by following the example of combining prompt words with Hires. fix, it is still possible to extract a few pictures.