wd14_tagger_inversion

## What's this

This is the dataset for inversing the embeddings of the given prediction result. We can get the embedding dataset with some tag combinations, and search images like this.

## How is this dataset made

This dataset is generated with anime images in danbooru (webp ones, based on [KBlueLeaf/danbooru2023-webp-4Mpixel](https://huggingface.co/datasets/KBlueLeaf/danbooru2023-webp-4Mpixel)), extracting the prediction results and the embeddings of each images.

## Model's goal

Train a model to inverse prediction result to embeddings. Maybe a MLP can be used? I don't know, just try them.

## 数据集简介
本数据集用于实现给定预测结果到嵌入向量的逆映射。通过指定标签组合即可获取对应嵌入数据集，并以此开展同类图像检索工作。

## 数据集构建流程
本数据集基于danbooru平台的动漫图像（webp格式，源自开源数据集[KBlueLeaf/danbooru2023-webp-4Mpixel](https://huggingface.co/datasets/KBlueLeaf/danbooru2023-webp-4Mpixel)）构建，通过提取每张图像的预测结果与嵌入向量生成。

## 模型训练目标
本任务的目标为训练模型完成从预测结果到嵌入向量的逆映射。可尝试使用多层感知机（MLP），具体实现方案尚未明确，可自由尝试各类可行方法。