Sign In

Useful Quality Embeddings XL

Updated: Apr 7, 2024
Oct 23, 2023
Base Model
SDXL 1.0
Trigger Words
default creator card background decoration
SDXL Training Contest Participant
Euge_'s Avatar

介绍(Chinese Ver.)

1. 简介

本模型用于 以 SDXL 为底的动漫模型,帮助优化所使用模型生成图像的:

  • 基础质量。包括增加 清晰度对比度 饱和度 至普遍认可的美学水平,(较强地)缓解严重内容崩坏扭曲杂乱

  • 人物肢体。包括改善人体结构面部 手部,效果因模型而有所差异;

我强烈鼓励您将它用于 任何 基于 SDXL 的动漫模型 之中,并于本页面的评论区给予反馈,包括:所使用的模型,具体效果。

模型使用方式与一般 embedding 模型相同。不同的是,您需要于 负面提示词 中触发该模型,而非正面提示词。

2. 实验

2023年10月29日更新:经过在 AIDXLv0.5 模型 [3] 上的测试,使用 v0.5 版本对图像的影响 几乎 是正面的。它在 绝大多数情况下 效果显著。

3. 制作方式

该模型的首个版本,使用文章 [1] 中提到的方法炼制,并与 XL 化后的 aid210 负面文本嵌入 [2] 融合后制成。这是一个实验性的产物,在 AIDXL [3] 系模型的表现良好。

负面 emb 的使用方法见 [2] 中的模型介绍。

对如何将基于 SD1.5 的 emb 模型转化为 SDXL 的 emb 模型有兴趣者可阅读 [1] 文章中的第四大节。

附录 A. Embedding 模型使用教程

如果您还未使用过 embedding 模型,请跟随以下使用教程:


安装 embedding 模型非常简单,只需将模型文件拖入:

  1. 对于 webui 使用者:{你的 webui 安装目录}/embeddings/ 文件夹内;

  2. 对于 ComfyUI 使用者: {你的 ComfyUI 安装目录}/models/embeddings/ 文件夹内;

  3. 对于其它 API 使用者,请参阅您使用 API 的说明。


通俗理解,embedding 模型是额外的自定义 tag,在 webui 中,其调用方式和触发方式与一般提示词无异,即,您可以通过在 webui 的提示词框内输入您 embedding 模型的文件名(不包含扩展名)触发。例如,假设您的 embedding 模型文件名为 "abc.safetensors",则您只需要在提示词框内输入 "abc" 即可触发。

用 ComfyUI 触发 embedding 则稍微不同,您需要在触发时,在 webui 的基础上为触发词添加前缀 "embedding:"。例如,假设您的 embedding 模型文件名为 "abc.safetensors",则您只需要在提示词框内输入 "embedding:abc" 即可触发。


Introduction (英文版)

1. Introduction

This model is used for animation models based on SDXL to help optimize the image generated by the model used. It can greatly improve:

  • Basic quality. This includes increasing clarity, contrast, and saturation to generally recognized aesthetic levels, and (strongly) mitigating severe content corruption, distortion, and clutter;

  • Human anatomy. Includes improvements to body anatomy, face and hands, with results varying by model;

I strongly encourage you to use it in any SDXL-based animation model, and give feedback in the comment area of this page, including: the model used and the specific effects.

The model is used in the same way as a normal embedding model. The difference is that you need to trigger the model on negative prompt area rather the positive.

2. Experiments

Update October 29, 2023: Tested on the AIDXLv0.5 model [3], the impact on the generated image of using the v0.5 version is almost positive. It works wonders in the vast majority of cases.

3. How to Train & Create

The negative text embedding, or neg emb for short, is refined using the method mentioned in article [1] and merged with the XL-ized aid210 negative text embedding [2]. This is an experimental product but performs well on the AIDXL [3] series models.

For the usage of negative embeddings, you may move to the model introduction in [2].

Those who are interested in how to convert an emb model based on SD1.5 into an emb model based on SDXL can read the fourth section of the article [1].

Appendix A. Tutorial for Embedding Model Usage

If you haven't used any embedding model yet, please follow the following tutorial:

Model Installation

To install an embedding model, you only need to move your model file into:

  1. For webui users: {your webui installation directory}/embeddings/ folder;

  2. For ComfyUI users: {your ComfyUI installation directory}/models/embeddings/ folder;

  3. For other API users, please see the instructions of your API.

Model Trigger

Simply put, the embedding model is just an additional custom tag. In webui, its triggering methods are the same as general prompt tags. That is, you can enter the filename of your embedding model (exclude its file extension) to trigger. For example, assuming your embedding model file is named "abc.safetensors". Then you only need to enter "abc" in the prompt area to trigger.

Triggering embedding with ComfyUI is slightly different. To do this, you need to add the prefix "embedding:" to the trigger word based on webui. For example, assuming your embedding model file is named "abc.safetensors". Then you need to enter "embedding:abc" in the prompt area to trigger.

For this model, the only difference is that you need to trigger within the negative prompt area instead of the positive.

引用 / References