Belle-Delphine

我要开发同款
匿名用户2024年07月31日
15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/Belle-Delphine
授权协议Apache License 2.0

作品详情

This is a Lora of the internet celebrity Belle Delphine.

There already exist multiple LoRAs and checkpoints of her, like here, here or here.

But I wasn’t happy with their results, as they either do not look how I would want it to look, or they are very prone to cat ears or pink hair. But everyone has their own taste, so check them out if you are unhappy with this one.

This LoRA does not use any trigger word, and I suggest a default weight of 1.

The preview images mainly use ChilloutMix, additional images are from Deliberate and Realistic Vision. The Overwatch images are using NeverEnding Dream.

The results are good enough for me to upload it here, but it is not perfect, and you might encounter the following problems:

  • It really likes generating the same lip shape, despite it not even being that frequent in the training images.

  • This also generates cat ears, especially if you have something with ear in your prompt or something on the head which could become ears, but it is less frequent than with other models in my experiments and opinion.

  • Depending on the prompt certain noise / “paint” on the face might appear.

Additionally, this LoRA is mostly doing its job only when the face is sufficiently big (also dependent on model). If it becomes too small, the quality decreases and I advise using the high resolution fix then. In the example images multiple images also make use of that.

This Lora is trained on the 1.5 base model with ~140 images, 50 repeats per image, 1e-5 learning rate, 5e-05 for the text encoder, 0.00015 for the unet. All training images were cropped to 512x512 and no bucketing was used.

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