metadata

task_categories:
  - text-to-image
language:
  - en
tags:
  - not-for-all-audiences
size_categories:
  - n<1K

Bang Dream 数据集 / Bang Dream Dataset

本数据集包含截止至 2025 年 1 月 31 日的《BanG Dream!》企划中全部 51 名角色的相关图片及元数据。 / This dataset contains images and metadata for all 51 characters in the "BanG Dream!" project as of January 31, 2025.

1. 数据集概述 / Dataset Overview

数据集主要包含以下几个部分： / The dataset mainly consists of the following parts:

角色信息 (bangdream_character_info.json) ：记录每个角色的详细信息，包括 ID、版权、核心标签、性别倾向、图片数量等。 / Character Information (bangdream_character_info.json): Records detailed information for each character, including ID, copyright, core tags, gender distribution, number of images, etc.
标签pid映射 (tag_mapping_full.json): 记录每个标签对应的图片 PID（图片唯一标识符）。 / Tag Mapping (tag_mapping_full.json): Records the PID (unique image identifier) of images corresponding to each tag.
角色图片：存储在 char_datasets/{hash_prefix}/{file_name} 目录下，包含不同类型的图片和数据文件。 / Character Images: Stored in the char_datasets/{hash_prefix}/{file_name} directory, containing different types of images and data files.

数据来源： 本数据集的图片及标签信息均来自 Danbooru 网站。 / Data Sources: The images and tag information in this dataset are from the Danbooru website.

2. 角色信息 (bangdream_character_info.json) / Character Information (bangdream_character_info.json)

bangdream_character_info.json 文件以 JSON 格式存储了每个角色的详细信息。示例如下： / The bangdream_character_info.json file stores detailed information for each character in JSON format. An example is as follows:

[
  {
    "_id": "mortis_(bang_dream!)",
    "copyright": {
      "bang_dream!": 1.0,
      "bang_dream!_it's_mygo!!!!!": 1.0
    },
    "core_tag_update_time": {
      "$date": "2025-02-21T05:21:42.583Z"
    },
    "core_tags": [
      "long_hair",
      "green_hair",
      "hat",
      "yellow_eyes",
      "red_hat",
      "beret",
      "ascot",
      "capelet"
    ],
    "gender": {
      "1girl": 0.3268633540372671
    },
    "id": 2258229,
    "post_count": 1288
  }
]

每个字段的含义如下： / The meaning of each field is as follows:

_id: 角色的唯一标识符, 对应 Danbooru 上的角色 tag。 / _id: Unique identifier for the character, corresponding to the character tag on Danbooru.
copyright: 版权信息，表示该角色在不同作品中的权重。 / copyright: Copyright information, indicating the weight of the character in different works.
core_tag_update_time: 核心标签的更新时间。 / core_tag_update_time: The update time of the core tags.
core_tags: 角色的核心标签列表。这些标签直接对应 Danbooru 上的角色 tag。 / core_tags: List of core tags for the character. These tags directly correspond to the character tags on Danbooru.
gender: 性别倾向。 / gender: Gender distribution.
id: 角色 ID。 / id: Character ID.
post_count: 包含该角色的图片数量。 / post_count: The number of images containing the character.

3. 标签映射 (tag_mapping_full) / Tag Mapping (tag_mapping_full)

tag_mapping_full 文件记录了每个标签对应的图片在Danbooru的PID。 / The tag_mapping_full file records the Danbooru PID of the images corresponding to each tag.

4. 图片存储结构 / Image Storage Structure

角色图片存储在 char_datasets 目录下，按照以下结构组织： / Character images are stored in the char_datasets directory, organized as follows:

char_datasets/
├── {hash_prefix}/
│   ├── {file_name}.tar
│   ├── {file_name}.npy
│   └── {file_name}.json

{hash_prefix}：根据角色名计算的哈希前缀。计算方法：将角色名（UTF-8 编码）进行 SHA256 哈希，取结果的前两位。 / {hash_prefix}: Hash prefix calculated from the character name. Calculation method: SHA256 hash the character name (UTF-8 encoded), and take the first two characters of the result.
{file_name}：经过处理后的角色名。 / {file_name}: The processed character name.

4.1 文件类型 / File Types

每个角色目录下包含三种类型的文件： / Each character directory contains three types of files:

{file_name}.tar: 包含角色图片的压缩包。
- train/1_char/: 原始图片 (WebP 格式)。
- train/1_stage/: 上半身和头像的碎片图 (WebP 格式)。
/ {file_name}.tar: Archive containing character images.
- train/1_char/: Original images (WebP format).
- train/1_stage/: Fragment images of the upper body and head (WebP format).
{file_name}.npy: 包含 CCIP 特征数据, 用于ccip模型识别 model name:ccip-caformer-24-randaug-pruned 。 / {file_name}.npy: Contains CCIP feature data, used for CCIP model recognition model name:ccip-caformer-24-randaug-pruned .
{file_name}.json: 包含 .tar 文件的元数据信息。 / {file_name}.json: Contains metadata information for the .tar file.

5. 文件名处理 / Filename Processing

角色名会经过以下步骤处理成文件名： / The character name is processed into a file name through the following steps:

规范化：移除非 ASCII 字符。使用 unicodedata.normalize('NFKD', filename).encode('ASCII', 'ignore').decode('ASCII') 实现。 / Normalization: Remove non-ASCII characters. Implemented using unicodedata.normalize('NFKD', filename).encode('ASCII', 'ignore').decode('ASCII').
替换非法字符：将 Windows 文件名中不允许的字符（<>:"/\\|?*）替换为下划线。使用 re.sub(r'[<>:"/\\|?*]', '_', filename) 实现。 / Replace Illegal Characters: Replace characters not allowed in Windows filenames (<>:"/\\|?*) with underscores. Implemented using re.sub(r'[<>:"/\\|?*]', '_', filename).
去除前后空格：使用 filename.strip()。 / Remove Leading/Trailing Spaces: Use filename.strip().
空文件名处理：如果文件名为空，则使用默认名称 "untitled"。 / Empty Filename Handling: If the filename is empty, use the default name "untitled".
长度截断：如果文件名长度超过 255 个字符，则截断至 255 个字符。 / Length Truncation: If the filename length exceeds 255 characters, truncate it to 255 characters.

Python 代码如下： / The Python code is as follows:

import unicodedata
import re

def sanitize_filename(filename):
    # 去除非 ASCII 字符
    # Remove non-ASCII characters
    filename = unicodedata.normalize('NFKD', filename).encode('ASCII', 'ignore').decode('ASCII')

    # 替换不允许的字符为下划线
    # Replace disallowed characters with underscores
    filename = re.sub(r'[<>:"/\\|?*]', '_', filename)

    # 去除前后空格
    # Remove leading and trailing whitespace
    filename = filename.strip()

    # 确保文件名不为空，如果为空则使用默认名称
    # Ensure the filename is not empty; use a default name if it is
    if not filename:
        filename = "untitled"

    # 截断文件名长度（如果需要）
    # Truncate the filename length (if necessary)
    max_length = 255  # 大多数文件系统的最大文件名长度
    # Maximum filename length for most file systems
    if len(filename) > max_length:
        filename = filename[:max_length]
    return filename