Datasets:
File size: 61,816 Bytes
bab67db |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 |
---
license: creativeml-openrail-m
---
# PureDanbooru Dataset

WarriorMama777/PureDanbooru
## 概要
このデータセットはDanbooruをメインに700万枚のイラストで構成された大規模な画像データセットです。
特徴は以下の通りです。
- Danbooruのタグを純粋に維持した未検閲のトレーニング用データセット。Danbooruのユーザーたちが地道に長年作業してきたタグ付けの作業がピュアに維持されています。
- 丁寧な前処理、およびトレーニング用にデータを整形済み。sd-scriptsのFinetuningにほぼ準拠した形でトレーニング用Jsonが提供されており、箱から出してすぐに使えます。
* * *
## Table of Contents
- [PureDanbooru Dataset](#puredanbooru-dataset)
- [概要](#概要)
- [Table of Contents](#table-of-contents)
- [データセットの詳細](#データセットの詳細)
- [画像の収集](#画像の収集)
- [内容](#内容)
- [ダウンロードコマンド例](#ダウンロードコマンド例)
- [前処理ツール](#前処理ツール)
- [サイズ](#サイズ)
- [画像の種類と形式および削減率](#画像の種類と形式および削減率)
- [タグとメタデータ](#タグとメタデータ)
- [メタデータ](#メタデータ)
- [タグ](#タグ)
- [形式](#形式)
- [変換処理](#変換処理)
- [キャプション](#キャプション)
- [json](#json)
- [相対パス](#相対パス)
- [相対パスでもデータセットの画像を読めるようにする](#相対パスでもデータセットの画像を読めるようにする)
- [絶対パス](#絶対パス)
- [変換処理サンプルコマンド](#変換処理サンプルコマンド)
- [編集方法](#編集方法)
- [使用方法](#使用方法)
- [ダウンロード](#ダウンロード)
- [すべてダウンロード](#すべてダウンロード)
- [特定のディレクトリをダウンロード](#特定のディレクトリをダウンロード)
- [解凍方法](#解凍方法)
- [使用例](#使用例)
- [データセットの構造](#データセットの構造)
- [ファイル構成](#ファイル構成)
- [データ形式](#データ形式)
- [ライセンス](#ライセンス)
- [引用](#引用)
- [謝辞](#謝辞)
- [連絡先情報](#連絡先情報)
- [更新履歴](#更新履歴)
- [免責事項](#免責事項)
- [-----English------](#-----english------)
- [PureDanbooru Dataset](#puredanbooru-dataset-1)
- [Overview](#overview)
- [Dataset Details](#dataset-details)
- [Image Collection](#image-collection)
- [Content](#content)
- [Download Command Example](#download-command-example)
- [Preprocessing Tool](#preprocessing-tool)
- [Size](#size)
- [Types and Formats of Images and Reduction Rate](#types-and-formats-of-images-and-reduction-rate)
- [Tags and Metadata](#tags-and-metadata)
- [Metadata](#metadata)
- [Tags](#tags)
- [Format](#format)
- [Conversion Process](#conversion-process)
- [Captions](#captions)
- [👁🗨VLM Test Result Sample (Anime illustration)](#vlm-test-result-sample-anime-illustration)
- [json](#json-1)
- [Relative Path](#relative-path)
- [Making dataset images readable with relative paths](#making-dataset-images-readable-with-relative-paths)
- [Absolute Path](#absolute-path)
- [Sample Command for Conversion Process](#sample-command-for-conversion-process)
- [Editing Method](#editing-method)
- [Usage](#usage)
- [Download](#download)
- [Download Everything](#download-everything)
- [Download Specific Directories](#download-specific-directories)
- [How to Decompress](#how-to-decompress)
- [Usage Example](#usage-example)
- [Dataset structure](#dataset-structure)
- [Directory structure](#directory-structure)
- [Data Format](#data-format)
- [License](#license)
- [Citation](#citation)
- [Acknowledgements](#acknowledgements)
- [Contact Information](#contact-information)
- [Update History](#update-history)
- [Disclaimer](#disclaimer)
* * *
## データセットの詳細
画像は主にDanbooruのイラストレーションで構成されています(Danbooruの一部のタグは隠されており有料会員しか閲覧できないので、その部分をGelbooru等から補填しています)。
### 画像の収集
主にDanbooruからgallery-dlを使用して収集したデータセットです。収集方法の詳細は以下の通りです。
- **収集方法**:
- Danbooruのクエリを用いて特定の検索ワードに対するURLを取得し、gallery-dlのためのコマンドリストを作成。
- 画像IDを15,000ごとに分割し、2024年4月時点でDanbooruに存在していたほぼすべての画像を数日から数週間かけて収集。
- **画像フィルタリング**:
- スコアがマイナスの画像を除外。
- 画像サイズが10KB以上15MB以下のものに限定。これは、非常に大きなキャンバスサイズの画像が収集作業の妨げになるのを防ぐためです。
- **注意点**:
- 収集過程でリクエストを大量に送信した結果、CloudflareからBANを受け、生IPではDanbooruから画像をダウンロードできなくなりました🥺 このため、VPNを使用して作業を継続しましたが、この方法は推奨しません。
#### 内容
- 初板
- Danbooru|id1 ~ id7440000
#### ダウンロードコマンド例
- Danbooruの検索クエリ:`score:>-1 id:1..15000 filetype:png,jpg,webp filesize:10kb..15m`
- gallery-dlのダウンロードコマンド例:`gallery-dl "https://danbooru.donmai.us/posts?tags=score%3A%3E-1+id%3A1..15000+filetype%3Apng%2Cjpg%2Cwebp+filesize%3A10kb..15M&z=5" -d "D:\dataset\danbooru\images_by_15000id" --write-metadata`
### 前処理ツール
データセットの規模と複雑性に対応するため、前処理スクリプトライブラリを同時に開発する必要がありました。このライブラリ"data-kitchen"は、すべてがAI によってプログラミングされたシンプルなコードで構成されていますが、データセットを構築する上での様々なタスクを効率的に処理することができました。
このカスタム前処理ライブラリのおかげで、データセットの構築プロセスが合理化され、大規模な画像コレクションを作成することができました。
[](https://github.com/WarriorMama777/data-kitchen)
### サイズ
- 画像/メタデータ/タグテキスト:それぞれ約700万
- 画像総ファイルサイズ:約1.3TB (圧縮時&解凍時)
- データセットjson:約2000万行、約40億文字数
### 画像の種類と形式および削減率
画像はすべて以下の形式に前処理されています。
- 画像形式:webp
- 品質:90
- 長辺が2048pxになるようにリサイズ
- 透過背景は白背景に変換
これによって、7.86 TB だった圧縮前のデータ量が、圧縮後には 1.3 TB になりました。これはデータ量を約 83.47% 削減したことになります。
### タグとメタデータ
#### メタデータ
メタデータはgallery-dlの`--write-metadata`を使ってDanbooruのオリジナルな情報を収集しています。
具体的には以下のような形式で保存されています。
```json
{
"id": 1,
"created_at": "2005-05-23T23:35:31.000-04:00",
"uploader_id": 1,
"score": 715,
"source": "http://www.biwa.ne.jp/~kyogoku/nekotama.jpg",
"md5": "d34e4cf0a437a5d65f8e82b7bcd02606",
"last_comment_bumped_at": "2024-03-11T16:09:34.477-04:00",
"rating": "s",
"image_width": 459,
"image_height": 650,
"tag_string": "1girl 2000s_(style) white_border white_thighhighs",
"fav_count": 1101,
"file_ext": "jpg",
"last_noted_at": null,
"parent_id": 57501,
"has_children": false,
"approver_id": null,
"tag_count_general": 61,
"tag_count_artist": 1,
"tag_count_character": 1,
"tag_count_copyright": 2,
"file_size": 127238,
"up_score": 702,
"down_score": -3,
"is_pending": false,
"is_flagged": false,
"is_deleted": false,
"tag_count": 69,
"updated_at": "2024-02-23T23:16:23.261-05:00",
"is_banned": false,
"pixiv_id": null,
"last_commented_at": "2024-05-23T13:44:46.633-04:00",
"has_active_children": false,
"bit_flags": 0,
"tag_count_meta": 4,
"has_large": false,
"has_visible_children": false,
"media_asset": {
"id": 1,
"created_at": "2005-05-23T23:35:31.000-04:00",
"updated_at": "2023-02-23T03:15:45.265-05:00",
"md5": "d34e4cf0a437a5d65f8e82b7bcd02606",
"file_ext": "jpg",
"file_size": 127238,
"image_width": 459,
"image_height": 650,
"duration": null,
"status": "active",
"file_key": "wUufKEivb",
"is_public": true,
"pixel_hash": "9c877dd5674d7fa251ce2de0c956fd36",
"variants": [
{
"type": "180x180",
"url": "https://cdn.donmai.us/180x180/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"width": 127,
"height": 180,
"file_ext": "jpg"
},
{
"type": "360x360",
"url": "https://cdn.donmai.us/360x360/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"width": 254,
"height": 360,
"file_ext": "jpg"
},
{
"type": "720x720",
"url": "https://cdn.donmai.us/720x720/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.webp",
"width": 459,
"height": 650,
"file_ext": "webp"
},
{
"type": "original",
"url": "https://cdn.donmai.us/original/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"width": 459,
"height": 650,
"file_ext": "jpg"
}
]
},
"tag_string_general": "1girl 2000s_(style) white_thighhighs",
"tag_string_character": "kousaka_tamaki",
"tag_string_copyright": "to_heart_(series) to_heart_2",
"tag_string_artist": "kyogoku_shin",
"tag_string_meta": "bad_id bad_link commentary translated",
"file_url": "https://cdn.donmai.us/original/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"large_file_url": "https://cdn.donmai.us/original/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"preview_file_url": "https://cdn.donmai.us/180x180/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"filename": "d34e4cf0a437a5d65f8e82b7bcd02606",
"extension": "jpg",
"date": "2005-05-24 03:35:31",
"tags": [
"1girl",
"white_thighhighs"
],
"tags_artist": [
"kyogoku_shin"
],
"tags_character": [
"kousaka_tamaki"
],
"tags_copyright": [
"to_heart_(series)",
"to_heart_2"
],
"tags_general": [
"1girl",
"2000s_(style)",
"white_thighhighs"
],
"tags_meta": [
"bad_id",
"bad_link",
"commentary",
"translated"
],
"search_tags": "score:>-1 id:1..15000 filetype:png,jpg,webp filesize:10kb..5M",
"category": "danbooru",
"subcategory": "tag"
}
```
#### タグ
##### 形式
タグはメタデータから以下のような形式に変換されています。
```txt
tenmaso,rozen_maiden,suiseiseki,illustration,|||,1girl,brown_hair,corset,dress,drill_hair,frills,green_dress,green_eyes,hat,heterochromia,lolita_fashion,long_hair,long_sleeves,red_eyes,ribbon,simple_background,solo,twin_drills,twintails,very_long_hair,|||,rating_g
```
タグは、下記のコマンドによってメタデータから以下のような形式に変換されています。
- Danbooruメタデータからプレーンテキストに変換された純粋なタグ
- すべてのタグに`illustration`を付与
- sd-scriptsで次のように指定することで使用できるセパレータ`|||`を、タグテキストの先頭と末尾に付与しています。:`--keep_tokens_separator="|||"`
- 並び順|「アーティスト名→作品名→キャラクター名→`illustration,|||`→タグテキスト→`|||`→レーティング」
##### 変換処理
タグはdata-kitchen/metadata_converter_danbooru.pyを使ってメタデータからプレーンテキストに変換され、tagフォルダに保存されています。変換処理のサンプルコマンドは以下のような形式になります。コマンドの具体的な内容については、data-kitchenリポジトリを確認してください。
```bash
python metadata_converter_danbooru.py --metadata_order "tag_string_artist" "tag_string_copyright" "tag_string_character" "tag_string_general" "rating" --dir "H:\dataset\metadata\Danbooru\images_by_15000id_01" --save_dir "H:\dataset\tag\Danbooru\images_by_15000id_01" --insert_custom_text 3 "`illustration,|||`" 4 "|||" --recursive --preserve_own_folder --preserve_structure --gc_disable --by_folder
```
#### キャプション
現状、次のような理由からキャプションは付属していません。
- ローカルで処理するにはコンピュータ等のリソースに限界がある
- 日本語性能の良いオープンソースVLMモデルが存在しない
CogVLMやMini-CPMは素晴らしいですが、ほとんど英語しか理解できず、多言語のOCRはできないようです。Gpt4oは本当に完璧ですが有料です。
どういうことか気になる人は、以下のトグルを確認してください🥺
<details><summary>👁🗨VLM Test Result Sample (Anime illustration)</summary>
▼Test01:OCR

プロンプト
> Please write a caption. It's a brief sentence that describes an image. Summarize it objectively and concisely, logically describing the main focus, actions, and context of the image. Start with the overall picture and then add specific details. Avoid personal opinions and non-essential details, and do not discuss the atmosphere of the scene or the emotions it may evoke. If there is text in the image, read it like OCR and describe it in the language it is written in (do not translate it into English unnecessarily). For background art, accurately describe objects and situations in the background. If there are no people, only describe the background. If there are characters, accurately describe features such as clothing and appearance. Summarize and describe this image in one paragraph. The base language for the description is English, but note that this does not apply to the OCR transcription function.
* * *
CogVLM2
> A young woman with short, dark hair and red eyes is captured in a moment of singing into a microphone. She wears a traditional kimono adorned with floral patterns, suggesting a cultural or celebratory event. The background is simple, with soft, warm colors and Japanese text, which may indicate the lyrics or a poetic context for her performance.
GLM 4v 9B
[cuuupid/glm-4v-9b – Run with an API on Replicate](https://replicate.com/cuuupid/glm-4v-9b)
> The image depicts an animated character with short, dark blue hair and red eyes holding a microphone, suggesting they are singing or speaking. The character is dressed in a traditional Japanese kimono with a floral pattern in shades of blue and purple. The background is plain white with Japanese text in a casual script, likely lyrics or a poem, to the right of the character
Gpt4o
> The image depicts an anime-style character with short dark blue hair and red eyes, dressed in a traditional blue floral kimono. The character is holding a microphone and appears to be singing. Surrounding the character, there is Japanese text: "勇気の鈴が りんりんりーん ♪ ふしぎな冒険 るんるんるーん ♪".
↓
> なんのキャラクターか分かりますか?
> ChatGPT
> はい、このキャラクターは『アイドルマスター シャイニーカラーズ』に登場する杜野凛世(もりの りんぜ)です。彼女は283プロダクションに所属するアイドルの一人で、静かでおしとやかな性格が特徴です。
* * *
▼Test02:キャラクター理解

CogVLM2:EN
> In a library setting, a young woman with short blue hair and glasses is engrossed in reading a green book. Beside her, a young woman with long brown hair tied in a ponytail leans in, her hand resting on the other woman's head, possibly in a gesture of comfort or to draw attention to something in the book.
Gpt4o
> The image depicts two characters from the anime "Sailor Moon." On the left is Makoto Kino (Sailor Jupiter), with long brown hair tied in a ponytail and green eyes, wearing a school uniform with a green collar and lacing detail. She is resting her head on one hand and using her other hand to pat the head of Ami Mizuno (Sailor Mercury), who has short blue hair and is wearing glasses. Ami is reading a green book and also dressed in a school uniform with a blue collar. The background features bookshelves filled with colorful books.
</details>
#### json
jsonフォルダには、タグと画像を結びつけたトレーニング用メタデータjsonファイルが保存されています。
##### 相対パス
相対パスで記載されたJson形式は以下のようになっています。
```json
{
"Danbooru\\images_by_15000id_01\\id_1__15000\\danbooru_1_d34e4cf0a437a5d65f8e82b7bcd02606.webp": {
"tags": "kyogoku_shin,to_heart_(series),to_heart_2,kousaka_tamaki,illustration,|||,1girl,2000s_(style),white_thighhighs,|||,rating_s"
},
"Danbooru\\images_by_15000id_01\\id_1__15000\\danbooru_2_710fd9cba4ef37260f9152ffa9d154d8.webp": {
"tags": "nakamura_hisashi,cardcaptor_sakura,comic_party,leaf_(studio),to_heart_(series),to_heart_2,trigun,card_master_peach,haga_reiko,kuhonbutsu_taishi,sendou_kazuki,takase_mizuki,yoshio_(leaf),yuzuhara_konomi,illustration,|||,2boys,2girls,ahoge,,|||,rating_s"
}
}
```
##### 相対パスでもデータセットの画像を読めるようにする
⚠sd-scriptsは現在、相対パスで記載されたJsonを読み込みません。以下のようにコードを変更する必要があります。
`train_util.py` スクリプトで相対パスに基づいてデータセットを正しく読み込めるように修正するには、次のように`train_util.py`スクリプトを修正します。具体的には、`abs_path` を設定する部分で、`os.path.exists` を使用する際に、相対パスが `subset.image_dir` からの相対パスであることを考慮して、そのパスを絶対パスに変換する処理を追加します。これによって、スクリプトがある場所に依存せずに、指定されたデータセットディレクトリ内のファイルを正しく参照できるようになります。
修正前のコード部分:
```python
if os.path.exists(image_key):
abs_path = image_key
else:
paths = glob_images(subset.image_dir, image_key)
if len(paths) > 0:
abs_path = paths[0]
```
修正後のコード:
```python
# 相対パスを絶対パスに変換するためのパスを結合
potential_path = os.path.join(subset.image_dir, image_key)
if os.path.exists(potential_path):
abs_path = potential_path
else:
# glob_imagesを使って、ディレクトリ内でマッチするファイルを探す
paths = glob_images(subset.image_dir, image_key)
if len(paths) > 0:
abs_path = paths[0]
```
この変更により、`image_key` が `subset.image_dir` からの相対パスである場合、そのパスを `subset.image_dir` に基づいて絶対パスに変換し、存在するかどうかをチェックします。`image_key` が直接存在しない場合に限り、`glob_images` 関数を使用してマッチするファイルを探します。これにより、入力された相対パスが機能し、データセットを正しく読み込むことができるようになります。
##### 絶対パス
現状、sd-scriptsは絶対パスであればコードを修正しなくても正しく画像を読み込むようです。
そのため絶対パスで記載されたJsonも同梱しています。形式は以下のようになっています。
```json
{
"H:\\Resources\\Resources_AI\\__Dataset\\forLeaningImages\\__Danbooru\\__BooruDiffusion\\dataset\\image\\Danbooru\\images_by_15000id_01\\id_1__15000\\danbooru_1_d34e4cf0a437a5d65f8e82b7bcd02606.webp": {
"tags": "kyogoku_shin,to_heart_(series),to_heart_2,kousaka_tamaki,illustration,|||,1girl,2000s_(style),white_thighhighs,|||,rating_s"
}
}
```
ただし記載されているパスはデータセット構築時のローカルのWindows環境下でのパスであり役に立ちませんので、適宜置換する必要があります。たとえばクラウド環境下(Runpod等)に置く場合は以下のように置換する必要があるでしょう
置換元:`H:\\Resources\\Resources_AI\\__Dataset\\forLeaningImages\\__Danbooru\\__BooruDiffusion\\dataset\\image`
置換先:`\\workspace\\__MyFolder\\image`
##### 変換処理サンプルコマンド
変換処理のサンプルコマンドは以下のような形式になります。コマンドの具体的な内容については、data-kitchenリポジトリを確認してください。
```bash
python metadata_merger.py --dir_base "H:\dataset\Danbooru\image\images_by_15000id_01\id_1__15000" --dir_append_data "H:\dataset\Danbooru\tag\images_by_15000id_01\id_1__15000" --dir_save_json "H:\dataset\Danbooru\json\datasets_metadata.json" --append_data_key tags --save_full_path
```
##### 編集方法
データセット用のjsonは非常に巨大で、一番大きいもので2000万の行、42億のテキストがあるため、編集する際はまともなテキストエディタで開いてください。
推奨:Notepad++
非推奨:VSCODE、EmEditor Free、メモ帳など
## 使用方法
### ダウンロード
ダウンロードはhuggingface-cliを推奨します。事前にhuggingface-cliが使えるようにインストールてください。
```bash
pip install huggingface-hub
```
#### すべてダウンロード
このデータセットをすべてダウンロードするには以下を実行します。
huggingfaceへのログインが求められた場合、huggingfaceのアクセストークンを入力してください。huggingfaceのアクセストークンを持っていない場合、huggingfaceでアクセストークンを作成してください。
```bash
huggingface-cli download WarriorMama777/PureDanbooru --repo-type dataset --local-dir "<download_directory_path>"
```
gitでも可能ですが、ダウンロード総量が肥大化するため非推奨です。
```bash
git lfs clone https://huggingface.co/WarriorMama777/PureDanbooru
```
#### 特定のディレクトリをダウンロード
特定のディレクトリをダウンロードするには、以下のコマンドを使用します:
```bash
huggingface-cli download WarriorMama777/PureDanbooru --repo-type dataset --local-dir "<download_directory_path>" --include <directory_to_download>
```
- `<repo_id>` はダウンロードしたいリポジトリのID(例:`username/model_name`)
- `<repo_type>` はリポジトリのタイプ(`model`、`dataset`、`space`のいずれか)
- `<download_directory_path>` はダウンロードしたファイルの保存先ディレクトリパス
- `<directory_to_download>` はリポジトリ内でダウンロードしたいディレクトリのパス
### 解凍方法
データセットはほとんどすべてがtarで圧縮されています。
tarを解凍する方法は様々ですが、以下にdata-kitchenの圧縮・解凍スクリプトを用いた解凍方法を記載しておきます。
コマンドの具体的な内容については、data-kitchenリポジトリを確認してください。
```bash
python packer.py --unpack --dir ".\image\Danbooru\images_by_15000id_01" --dir_save ".\image\Danbooru\images_by_15000id_01" --by_pack
```
### 使用例
実際にsd-scriptsを使用して学習する場合、以下のようになるでしょう。
以下は`workspace`に`__MyFolder`というディレクトリを作成し、そこにPureDanbooruをダウンロード、解凍した場合を想定しています。
ただしtomlを使用して行う方法もありますので、その場合は各自でお調べください。
```bash
accelerate launch --num_processes=4 --multi_gpu --num_machines=1 --gpu_ids=0,1,2,3 sdxl_train.py --pretrained_model_name_or_path "/workspace/stable-diffusion-webui/models/Stable-diffusion/sdxl.safetensors" --in_json "/workspace/__MyFolder/json/datasets_metadata_danbooru_images_by_15000id_01-02_hiddentag_relative-path.json" --train_data_dir "/workspace/__MyFolder/image" --output_dir "/workspace/__MyFolder/model" --precision "float" --save_precision "float" --mixed_precision "fp16" --xformers --gradient_checkpointing --enable_bucket --save_model_as "safetensors" --resolution 1024 --train_batch_size 16 --max_train_epochs 20 --save_every_n_epochs 2 --optimizer_type "prodigy" --learning_rate 1 --lr_scheduler_num_cycles 5 --lr_scheduler "cosine" --output_name "Finetuning_SDXL_PureDanbooru_v1_prodigy" --max_token_length 225 --shuffle_caption --seed 42 --logging_dir "/workspace/__MyFolder/log" --log_prefix=Finetuning_SDXL_PureDanbooru_v1_prodigy --fp8_base --optimizer_args "d_coef=0.8" "decouple=True" "weight_decay=0.01" "use_bias_correction=True" "safeguard_warmup=True" "betas=0.9,0.99" "eps=1e-08" --keep_tokens_separator="|||" --noise_offset 0.05 --noise_offset_random_strength
```
## データセットの構造
### ファイル構成
.
├── alphachannel
│ └── Danbooru
│ ├── images_by_15000id_01
│ │ ├── id_1__15000.tar
│ │ ├── ~~~~~~~~~~~~
│ │ └── id_3330001__3345000.tar
│ └── images_by_15000id_02
│ ├── id_3345001__3360000.tar
│ ├── ~~~~~~~~~~~~
│ └── id_7425001__7440000.tar
├── duplicate_image
│ ├── Danbooru
│ │ ├── images_by_15000id_01
│ │ │ ├── id_1__15000.tar
│ │ │ ├── ~~~~~~~~~~~~
│ │ │ └── id_3330001__3345000.tar
│ │ └── images_by_15000id_02
│ │ ├── id_3345001__3360000.tar
│ │ ├── ~~~~~~~~~~~~
│ │ └── id_7425001__7440000.tar
│ └── Gelbooru
│ ├── hidden_tag_01.tar
│ ├── hidden_tag_02.tar
│ └── hidden_tag_03.tar
├── image
│ ├── Danbooru
│ │ ├── images_by_15000id_01
│ │ │ ├── id_1__15000.tar
│ │ │ ├── ~~~~~~~~~~~~
│ │ │ └── id_3330001__3345000.tar
│ │ └── images_by_15000id_02
│ │ ├── id_3345001__3360000.tar
│ │ ├── ~~~~~~~~~~~~
│ │ └── id_7425001__7440000.tar
│ └── Gelbooru
├── json
│ ├── datasets_metadata_danbooru_images_by_15000id_01-02_full-path.json
│ ├── datasets_metadata_danbooru_images_by_15000id_01-02_hiddentag_full-path.json
│ ├── datasets_metadata_danbooru_images_by_15000id_01-02_hiddentag_relative-path.json
│ └── datasets_metadata_danbooru_images_by_15000id_01-02_relative-path.json
├── metadata
│ ├── Danbooru
│ │ ├── images_by_15000id_01
│ │ │ ├── id_1__15000.tar
│ │ │ ├── ~~~~~~~~~~~~
│ │ │ └── id_3330001__3345000.tar
│ │ └── images_by_15000id_02
│ │ ├── id_3345001__3360000.tar
│ │ ├──~~~~~~~~~~~~
│ │ └── id_7425001__7440000.tar
│ └── Gelbooru
│ ├── hidden_tag_01.tar
│ ├── hidden_tag_02.tar
│ └── hidden_tag_03.tar
└── tag
├── Danbooru
│ ├── images_by_15000id_01
│ │ ├── id_1__15000.tar
│ │ ├──~~~~~~~~~~~~
│ │ └── id_3330001__3345000.tar
│ └── images_by_15000id_02
│ ├── id_3345001__3360000.tar
│ ├── ~~~~~~~~~~~~
│ └── id_7425001__7440000.tar
└── Gelbooru
├── hidden_tag_01.tar
├── hidden_tag_02.tar
└── hidden_tag_03.tar
### データ形式
これはtarで固めたデータセットをアップロードしただけの素朴なデータセットです。
SQL🥃? Datasets🤗?? Dask🍡??? Croissant🥐??????? Parquet🤯?????????
なにもわからない。
## ライセンス
creativeml-openrail-m
## 引用
強制ではありませんが、引用する場合は以下の形式を使用すると良いでしょう。
```
WarriorMama777. (2024). "PureDanbooru" (Version 1.0). Hugging Face. https://huggingface.co/datasets/WarriorMama777/PureDanbooru
```
## 謝辞
Danbooruおよび、Danbooruで長年にわたって地道にタグ付け作業してきたユーザーの仕事に感謝します。アーティストの皆様に対しては一方的に謝辞を送ることは控えますが、それぞれの仕事に対して敬意を表します。
## 連絡先情報
当huggingfaceリポジトリのディスカッション欄へどうぞ。
## 更新履歴
| バージョン | 日付 | 変更内容 |
| --- | --- | --- |
| v1.0.0 | 2024-07-07 | 初板 |
## 免責事項
<details><summary>READ MORE: Disclaimer (JPN)</summary> 1. AIの学習等における機械学習時の著作物の利用については、日本国著作権法第三十条の四第二項「情報解析(多数の著作物その他の大量の情報から、当該情報を構成する言語、音、影像その他の要素に係る情報を抽出し、比較、分類その他の解析を行うことをいう。第四十七条の五第一項第二号において同じ。)の用に供する場合」を参照し行っております。 [著作権法 | e-Gov法令検索: デジタル庁](https://elaws.e-gov.go.jp/document?lawid=345AC0000000048#Mp-At_30) 2. 著作物の掲載にあたっては、私的使用や研究目的である場合、商品化の資料として使用する場合(但し、商品化されるものが著作物の主たる利用目的となる場合にはこの限りではない)、批評・論評・報道での利用である場合、パロディや二次創作としてオリジナリティを発揮する場合等の例外的な著作物の使用が認められるケースに該当すると考え掲載しています。 3. 当該学習データセットは、わいせつな視覚コンテンツを直ちに表示されないよう配慮されており、学習データセットによる投稿が公共の場においてわいせつな内容を表示することを意図するものではありません。 なお、著作権者の意に反する利用や不当な利益を得る利用は避けるべきであり、著作権保有者によって正当な理由をもって当該データセットの関係する過半数の関係者から申し立てされた場合、掲載は速やかに中止されることが保証されます。 また本人は、使用する機械学習データセット中に、著作権者の権利を十分に尊重していないと評価され得る可能性を持つ情報源からのデータが含まれ得ることを認識しており、開発者等関係者が依拠すべき法律の整備等の動向について適切な監視と管理の義務を負っています。 本人は、法律や倫理に反する行為を意図しておらず、著作権者を含むイラストレーター及びアーティスト等の権利を尊重し、人工知能の開発を進めることに努めます。</details>
# -----English------
# PureDanbooru Dataset

WarriorMama777/PureDanbooru
## Overview
This dataset is a large-scale image dataset composed of 7 million illustrations, primarily from Danbooru.
Its features are as follows:
- An uncensored training dataset that purely maintains Danbooru tags. The tagging work that Danbooru users have diligently carried out over many years is purely maintained.
- Carefully preprocessed and formatted data for training. Training JSON is provided in a format compliant with sd-scripts Finetuning, ready to use out of the box.
## Dataset Details
The images are mainly composed of illustrations from Danbooru (some tags on Danbooru are hidden and can only be viewed by paid members, so those parts are supplemented from sources like Gelbooru).
### Image Collection
This dataset was mainly collected from Danbooru using gallery-dl. The details of the collection method are as follows:
- **Collection Method**:
- Created a command list for gallery-dl by obtaining URLs for specific search words using Danbooru queries.
- Divided image IDs into groups of 15,000 and collected almost all images that existed on Danbooru as of April 2024 over a period of several days to several weeks.
- **Image Filtering**:
- Excluded images with negative scores.
- Limited to images with file sizes between 10KB and 15MB. This is to prevent very large canvas size images from hindering the collection process.
- **Note**:
- As a result of sending a large number of requests during the collection process, I was banned by Cloudflare and could no longer download images from Danbooru using a raw IP 🥺 Therefore, I continued my work using a VPN, but this method is not recommended.
#### Content
- Initial version
- Danbooru | id1 ~ id7440000
#### Download Command Example
- Danbooru search query: `score:>-1 id:1..15000 filetype:png,jpg,webp filesize:10kb..15m`
- gallery-dl download command example: `gallery-dl "https://danbooru.donmai.us/posts?tags=score%3A%3E-1+id%3A1..15000+filetype%3Apng%2Cjpg%2Cwebp+filesize%3A10kb..15M&z=5" -d "D:\dataset\danbooru\images_by_15000id" --write-metadata`
### Preprocessing Tool
To handle the scale and complexity of the dataset, I needed to simultaneously develop a preprocessing script library. This library "data-kitchen" consists entirely of simple code programmed by AI, but it was able to efficiently process various tasks in building the dataset.
Thanks to this custom preprocessing library, the dataset construction process was streamlined, allowing me to create a large-scale image collection.
[](https://github.com/WarriorMama777/data-kitchen)

WarriorMama777/data-kitchen
### Size
- Images/Metadata/Tag text: About 7 million each
- Total image file size: About 1.3TB (when compressed & uncompressed)
- Dataset json: About 20 million lines, about 4 billion characters
### Types and Formats of Images and Reduction Rate
All images have been preprocessed to the following format:
- Image format: webp
- Quality: 90
- Resized so that the long side becomes 2048px
- Transparent backgrounds converted to white backgrounds
This reduced the pre-compression data volume of 7.86 TB to 1.3 TB after compression. This represents a data volume reduction of about 83.47%.
### Tags and Metadata
#### Metadata
Metadata is collected using gallery-dl's `--write-metadata` to gather original information from Danbooru.
Specifically, it is saved in the following format:
```json
{
"id": 1,
"created_at": "2005-05-23T23:35:31.000-04:00",
"uploader_id": 1,
"score": 715,
"source": "http://www.biwa.ne.jp/~kyogoku/nekotama.jpg",
"md5": "d34e4cf0a437a5d65f8e82b7bcd02606",
"last_comment_bumped_at": "2024-03-11T16:09:34.477-04:00",
"rating": "s",
"image_width": 459,
"image_height": 650,
"tag_string": "1girl 2000s_(style) white_border white_thighhighs",
"fav_count": 1101,
"file_ext": "jpg",
"last_noted_at": null,
"parent_id": 57501,
"has_children": false,
"approver_id": null,
"tag_count_general": 61,
"tag_count_artist": 1,
"tag_count_character": 1,
"tag_count_copyright": 2,
"file_size": 127238,
"up_score": 702,
"down_score": -3,
"is_pending": false,
"is_flagged": false,
"is_deleted": false,
"tag_count": 69,
"updated_at": "2024-02-23T23:16:23.261-05:00",
"is_banned": false,
"pixiv_id": null,
"last_commented_at": "2024-05-23T13:44:46.633-04:00",
"has_active_children": false,
"bit_flags": 0,
"tag_count_meta": 4,
"has_large": false,
"has_visible_children": false,
"media_asset": {
"id": 1,
"created_at": "2005-05-23T23:35:31.000-04:00",
"updated_at": "2023-02-23T03:15:45.265-05:00",
"md5": "d34e4cf0a437a5d65f8e82b7bcd02606",
"file_ext": "jpg",
"file_size": 127238,
"image_width": 459,
"image_height": 650,
"duration": null,
"status": "active",
"file_key": "wUufKEivb",
"is_public": true,
"pixel_hash": "9c877dd5674d7fa251ce2de0c956fd36",
"variants": [
{
"type": "180x180",
"url": "https://cdn.donmai.us/180x180/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"width": 127,
"height": 180,
"file_ext": "jpg"
},
{
"type": "360x360",
"url": "https://cdn.donmai.us/360x360/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"width": 254,
"height": 360,
"file_ext": "jpg"
},
{
"type": "720x720",
"url": "https://cdn.donmai.us/720x720/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.webp",
"width": 459,
"height": 650,
"file_ext": "webp"
},
{
"type": "original",
"url": "https://cdn.donmai.us/original/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"width": 459,
"height": 650,
"file_ext": "jpg"
}
]
},
"tag_string_general": "1girl 2000s_(style) white_thighhighs",
"tag_string_character": "kousaka_tamaki",
"tag_string_copyright": "to_heart_(series) to_heart_2",
"tag_string_artist": "kyogoku_shin",
"tag_string_meta": "bad_id bad_link commentary translated",
"file_url": "https://cdn.donmai.us/original/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"large_file_url": "https://cdn.donmai.us/original/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"preview_file_url": "https://cdn.donmai.us/180x180/d3/4e/d34e4cf0a437a5d65f8e82b7bcd02606.jpg",
"filename": "d34e4cf0a437a5d65f8e82b7bcd02606",
"extension": "jpg",
"date": "2005-05-24 03:35:31",
"tags": [
"1girl",
"white_thighhighs"
],
"tags_artist": [
"kyogoku_shin"
],
"tags_character": [
"kousaka_tamaki"
],
"tags_copyright": [
"to_heart_(series)",
"to_heart_2"
],
"tags_general": [
"1girl",
"2000s_(style)",
"white_thighhighs"
],
"tags_meta": [
"bad_id",
"bad_link",
"commentary",
"translated"
],
"search_tags": "score:>-1 id:1..15000 filetype:png,jpg,webp filesize:10kb..5M",
"category": "danbooru",
"subcategory": "tag"
}
```
#### Tags
##### Format
Tags are converted from metadata to the following format:
```txt
tenmaso,rozen_maiden,suiseiseki,illustration,|||,1girl,brown_hair,corset,dress,drill_hair,frills,green_dress,green_eyes,hat,heterochromia,lolita_fashion,long_hair,long_sleeves,red_eyes,ribbon,simple_background,solo,twin_drills,twintails,very_long_hair,|||,rating_g
```
Tags are converted from metadata to the following format using these commands:
- Pure tags converted from Danbooru metadata to plain text
- `illustration` is added to all tags
- The separator `|||` is added to the beginning and end of the tag text, which can be used by specifying it in sd-scripts as follows: `--keep_tokens_separator="|||"`
- Order | "Artist name → Work name → Character name → `illustration,|||` → Tag text → `|||` → Rating"
##### Conversion Process
Tags are converted from metadata to plain text using data-kitchen/metadata_converter_danbooru.py and saved in the tag folder. A sample command for the conversion process is in the following format. For specific details about the command, please check the data-kitchen repository.
```bash
python metadata_converter_danbooru.py --metadata_order "tag_string_artist" "tag_string_copyright" "tag_string_character" "tag_string_general" "rating" --dir "H:\dataset\metadata\Danbooru\images_by_15000id_01" --save_dir "H:\dataset\tag\Danbooru\images_by_15000id_01" --insert_custom_text 3 "`illustration,|||`" 4 "|||" --recursive --preserve_own_folder --preserve_structure --gc_disable --by_folder
```
#### Captions
Currently, captions are not included for the following reasons:
- There are limitations in computer and other resources for local processing
- There are no open-source VLM models with good Japanese performance
CogVLM and Mini-CPM are excellent, but they seem to understand almost only English and can't perform multilingual OCR. GPT4o is truly perfect but it's paid.
If you want to know what this means, check it out below 🥺
##### 👁🗨VLM Test Result Sample (Anime illustration)
▼Test01: OCR

Prompt
> Please write a caption. It's a brief sentence that describes an image. Summarize it objectively and concisely, logically describing the main focus, actions, and context of the image. Start with the overall picture and then add specific details. Avoid personal opinions and non-essential details, and do not discuss the atmosphere of the scene or the emotions it may evoke. If there is text in the image, read it like OCR and describe it in the language it is written in (do not translate it into English unnecessarily). For background art, accurately describe objects and situations in the background. If there are no people, only describe the background. If there are characters, accurately describe features such as clothing and appearance. Summarize and describe this image in one paragraph. The base language for the description is English, but note that this does not apply to the OCR transcription function.
* * *
CogVLM2
> A young woman with short, dark hair and red eyes is captured in a moment of singing into a microphone. She wears a traditional kimono adorned with floral patterns, suggesting a cultural or celebratory event. The background is simple, with soft, warm colors and Japanese text, which may indicate the lyrics or a poetic context for her performance.
GLM 4v 9B
[cuuupid/glm-4v-9b – Run with an API on Replicate](https://replicate.com/cuuupid/glm-4v-9b)
> The image depicts an animated character with short, dark blue hair and red eyes holding a microphone, suggesting they are singing or speaking. The character is dressed in a traditional Japanese kimono with a floral pattern in shades of blue and purple. The background is plain white with Japanese text in a casual script, likely lyrics or a poem, to the right of the character
Gpt4o
> The image depicts an anime-style character with short dark blue hair and red eyes, dressed in a traditional blue floral kimono. The character is holding a microphone and appears to be singing. Surrounding the character, there is Japanese text: "勇気の鈴が りんりんりーん ♪ ふしぎな冒険 るんるんるーん ♪".
↓
> なんのキャラクターか分かりますか?
> ChatGPT
> はい、このキャラクターは『アイドルマスター シャイニーカラーズ』に登場する杜野凛世(もりの りんぜ)です。彼女は283プロダクションに所属するアイドルの一人で、静かでおしとやかな性格が特徴です。
* * *
▼Test02: Understanding the characters

CogVLM2:EN
> In a library setting, a young woman with short blue hair and glasses is engrossed in reading a green book. Beside her, a young woman with long brown hair tied in a ponytail leans in, her hand resting on the other woman's head, possibly in a gesture of comfort or to draw attention to something in the book.
Gpt4o
> The image depicts two characters from the anime "Sailor Moon." On the left is Makoto Kino (Sailor Jupiter), with long brown hair tied in a ponytail and green eyes, wearing a school uniform with a green collar and lacing detail. She is resting her head on one hand and using her other hand to pat the head of Ami Mizuno (Sailor Mercury), who has short blue hair and is wearing glasses. Ami is reading a green book and also dressed in a school uniform with a blue collar. The background features bookshelves filled with colorful books.
#### json
The json folder contains training metadata json files that link tags and images.
##### Relative Path
The Json format with relative paths is as follows:
```json
{
"Danbooru\\images_by_15000id_01\\id_1__15000\\danbooru_1_d34e4cf0a437a5d65f8e82b7bcd02606.webp": {
"tags": "kyogoku_shin,to_heart_(series),to_heart_2,kousaka_tamaki,illustration,|||,1girl,2000s_(style),white_thighhighs,|||,rating_s"
},
"Danbooru\\images_by_15000id_01\\id_1__15000\\danbooru_2_710fd9cba4ef37260f9152ffa9d154d8.webp": {
"tags": "nakamura_hisashi,cardcaptor_sakura,comic_party,leaf_(studio),to_heart_(series),to_heart_2,trigun,card_master_peach,haga_reiko,kuhonbutsu_taishi,sendou_kazuki,takase_mizuki,yoshio_(leaf),yuzuhara_konomi,illustration,|||,2boys,2girls,ahoge,,|||,rating_s"
}
}
```
##### Making dataset images readable with relative paths
⚠ Currently, sd-scripts does not read Json written with relative paths. The code needs to be modified as follows:
To modify the `train_util.py` script to correctly read the dataset based on relative paths, you need to modify the `train_util.py` script as follows. Specifically, when setting `abs_path`, add a process to convert the relative path to an absolute path, considering that the relative path is relative to `subset.image_dir` when using `os.path.exists`. This allows the script to correctly reference files in the specified dataset directory without depending on where the script is located.
Before modification:
```python
if os.path.exists(image_key):
abs_path = image_key
else:
paths = glob_images(subset.image_dir, image_key)
if len(paths) > 0:
abs_path = paths[0]
```
After modification:
```python
# Combine paths to convert relative path to absolute path
potential_path = os.path.join(subset.image_dir, image_key)
if os.path.exists(potential_path):
abs_path = potential_path
else:
# Use glob_images to find matching files in the directory
paths = glob_images(subset.image_dir, image_key)
if len(paths) > 0:
abs_path = paths[0]
```
With this change, if `image_key` is a relative path from `subset.image_dir`, it converts the path to an absolute path based on `subset.image_dir` and checks if it exists. Only if `image_key` doesn't exist directly, it uses the `glob_images` function to find matching files. This allows the input relative paths to function and correctly load the dataset.
##### Absolute Path
Currently, it seems that sd-scripts can correctly read images with absolute paths without modifying the code.
Therefore, I've also included Json written with absolute paths. The format is as follows:
```json
{
"H:\\Resources\\Resources_AI\\__Dataset\\forLeaningImages\\__Danbooru\\__BooruDiffusion\\dataset\\image\\Danbooru\\images_by_15000id_01\\id_1__15000\\danbooru_1_d34e4cf0a437a5d65f8e82b7bcd02606.webp": {
"tags": "kyogoku_shin,to_heart_(series),to_heart_2,kousaka_tamaki,illustration,|||,1girl,2000s_(style),white_thighhighs,|||,rating_s"
}
}
```
However, the paths listed are from the local Windows environment at the time of dataset construction and are not useful, so they need to be replaced as appropriate. For example, if you're placing it in a cloud environment (like Runpod), you'll need to replace it as follows:
Replace from: `H:\\Resources\\Resources_AI\\__Dataset\\forLeaningImages\\__Danbooru\\__BooruDiffusion\\dataset\\image`
Replace to: `\\workspace\\__MyFolder\\image`
##### Sample Command for Conversion Process
The sample command for the conversion process is in the following format. For specific details about the command, please check the data-kitchen repository.
```bash
python metadata_merger.py --dir_base "H:\dataset\Danbooru\image\images_by_15000id_01\id_1__15000" --dir_append_data "H:\dataset\Danbooru\tag\images_by_15000id_01\id_1__15000" --dir_save_json "H:\dataset\Danbooru\json\datasets_metadata.json" --append_data_key tags --save_full_path
```
##### Editing Method
The json for the dataset is extremely large, with the largest having 20 million lines and 4.2 billion characters of text, so please use a decent text editor when editing.
Recommended: Notepad++
Not recommended: VSCODE, EmEditor Free, Notepad, etc.
## Usage
### Download
I recommend using huggingface-cli for downloading. Please install huggingface-cli in advance so that it can be used.
```bash
pip install huggingface-hub
```
#### Download Everything
To download this entire dataset, execute the following:
If you're asked to log in to huggingface, enter your huggingface access token. If you don't have a huggingface access token, create one on huggingface.
```bash
huggingface-cli download WarriorMama777/PureDanbooru --repo-type dataset --local-dir "<download_directory_path>"
```
It's also possible with git, but not recommended as the total download size becomes bloated.
```bash
git lfs clone https://huggingface.co/WarriorMama777/PureDanbooru
```
#### Download Specific Directories
To download specific directories, use the following command:
```bash
huggingface-cli download WarriorMama777/PureDanbooru --repo-type dataset --local-dir "<download_directory_path>" --include <directory_to_download>
```
- `<repo_id>` is the ID of the repository you want to download (e.g., `username/model_name`)
- `<repo_type>` is the type of repository (`model`, `dataset`, or `space`)
- `<download_directory_path>` is the directory path where you want to save the downloaded files
- `<directory_to_download>` is the path of the directory you want to download within the repository
### How to Decompress
Almost all of the dataset is compressed with tar.
There are various ways to decompress tar, but I'll describe how to decompress using the compression/decompression script from data-kitchen.
For specific details about the command, please check the data-kitchen repository.
```bash
python packer.py --unpack --dir ".\image\Danbooru\images_by_15000id_01" --dir_save ".\image\Danbooru\images_by_15000id_01" --by_pack
```
### Usage Example
When actually using sd-scripts for training, it would be like this:
The following assumes that you've created a `__MyFolder` directory in `workspace` and downloaded and decompressed PureDanbooru there.
However, there's also a method using toml, so please research that on your own if you prefer.
```bash
accelerate launch --num_processes=4 --multi_gpu --num_machines=1 --gpu_ids=0,1,2,3 sdxl_train.py --pretrained_model_name_or_path "/workspace/stable-diffusion-webui/models/Stable-diffusion/sdxl.safetensors" --in_json "/workspace/__MyFolder/json/datasets_metadata_danbooru_images_by_15000id_01-02_hiddentag_relative-path.json" --train_data_dir "/workspace/__MyFolder/image" --output_dir "/workspace/__MyFolder/model" --precision "float" --save_precision "float" --mixed_precision "fp16" --xformers --gradient_checkpointing --enable_bucket --save_model_as "safetensors" --resolution 1024 --train_batch_size 16 --max_train_epochs 20 --save_every_n_epochs 2 --optimizer_type "prodigy" --learning_rate 1 --lr_scheduler_num_cycles 5 --lr_scheduler "cosine" --output_name "Finetuning_SDXL_PureDanbooru_v1_prodigy" --max_token_length 225 --shuffle_caption --seed 42 --logging_dir "/workspace/__MyFolder/log" --log_prefix=Finetuning_SDXL_PureDanbooru_v1_prodigy --fp8_base --optimizer_args "d_coef=0.8" "decouple=True" "weight_decay=0.01" "use_bias_correction=True" "safeguard_warmup=True" "betas=0.9,0.99" "eps=1e-08" --keep_tokens_separator="|||" --noise_offset 0.05 --noise_offset_random_strength
```
## Dataset structure
### Directory structure
.
├── alphachannel
│ └── Danbooru
│ ├── images_by_15000id_01
│ │ ├── id_1__15000.tar
│ │ ├── ~~~~~~~~~~~~
│ │ └── id_3330001__3345000.tar
│ └── images_by_15000id_02
│ ├── id_3345001__3360000.tar
│ ├── ~~~~~~~~~~~~
│ └── id_7425001__7440000.tar
├── duplicate_image
│ ├── Danbooru
│ │ ├── images_by_15000id_01
│ │ │ ├── id_1__15000.tar
│ │ │ ├── ~~~~~~~~~~~~
│ │ │ └── id_3330001__3345000.tar
│ │ └── images_by_15000id_02
│ │ ├── id_3345001__3360000.tar
│ │ ├── ~~~~~~~~~~~~
│ │ └── id_7425001__7440000.tar
│ └── Gelbooru
│ ├── hidden_tag_01.tar
│ ├── hidden_tag_02.tar
│ └── hidden_tag_03.tar
├── image
│ ├── Danbooru
│ │ ├── images_by_15000id_01
│ │ │ ├── id_1__15000.tar
│ │ │ ├── ~~~~~~~~~~~~
│ │ │ └── id_3330001__3345000.tar
│ │ └── images_by_15000id_02
│ │ ├── id_3345001__3360000.tar
│ │ ├── ~~~~~~~~~~~~
│ │ └── id_7425001__7440000.tar
│ └── Gelbooru
├── json
│ ├── datasets_metadata_danbooru_images_by_15000id_01-02_full-path.json
│ ├── datasets_metadata_danbooru_images_by_15000id_01-02_hiddentag_full-path.json
│ ├── datasets_metadata_danbooru_images_by_15000id_01-02_hiddentag_relative-path.json
│ └── datasets_metadata_danbooru_images_by_15000id_01-02_relative-path.json
├── metadata
│ ├── Danbooru
│ │ ├── images_by_15000id_01
│ │ │ ├── id_1__15000.tar
│ │ │ ├── ~~~~~~~~~~~~
│ │ │ └── id_3330001__3345000.tar
│ │ └── images_by_15000id_02
│ │ ├── id_3345001__3360000.tar
│ │ ├──~~~~~~~~~~~~
│ │ └── id_7425001__7440000.tar
│ └── Gelbooru
│ ├── hidden_tag_01.tar
│ ├── hidden_tag_02.tar
│ └── hidden_tag_03.tar
└── tag
├── Danbooru
│ ├── images_by_15000id_01
│ │ ├── id_1__15000.tar
│ │ ├──~~~~~~~~~~~~
│ │ └── id_3330001__3345000.tar
│ └── images_by_15000id_02
│ ├── id_3345001__3360000.tar
│ ├── ~~~~~~~~~~~~
│ └── id_7425001__7440000.tar
└── Gelbooru
├── hidden_tag_01.tar
├── hidden_tag_02.tar
└── hidden_tag_03.tar
### Data Format
This is a simple dataset that just uploads tar-compressed data.
SQL🥃? Datasets🤗?? Dask🍡??? Croissant🥐??????? Parquet🤯?????????
I don't understand any of it.
## License
creativeml-openrail-m
## Citation
While not mandatory, if you wish to cite this dataset, you could use the following format:
```
WarriorMama777. (2024). "PureDanbooru" (Version 1.0). Hugging Face. https://huggingface.co/datasets/WarriorMama777/PureDanbooru
```
## Acknowledgements
I'm grateful to Danbooru and the users who have diligently tagged on Danbooru over the years. I refrain from unilaterally sending thanks to the artists, but I respect each of their works.
## Contact Information
Please use the discussion section of this huggingface repository.
## Update History
| Version | Date | Changes |
| --- | --- | --- |
| v1.0.0 | 2024-07-07 | Initial release |
## Disclaimer
※To prevent errors through translation, it is posted in the original text.
<details><summary>READ MORE: Disclaimer (JPN)</summary> 1. AIの学習等における機械学習時の著作物の利用については、日本国著作権法第三十条の四第二項「情報解析(多数の著作物その他の大量の情報から、当該情報を構成する言語、音、影像その他の要素に係る情報を抽出し、比較、分類その他の解析を行うことをいう。第四十七条の五第一項第二号において同じ。)の用に供する場合」を参照し行っております。 [著作権法 | e-Gov法令検索: デジタル庁](https://elaws.e-gov.go.jp/document?lawid=345AC0000000048#Mp-At_30) 2. 著作物の掲載にあたっては、私的使用や研究目的である場合、商品化の資料として使用する場合(但し、商品化されるものが著作物の主たる利用目的となる場合にはこの限りではない)、批評・論評・報道での利用である場合、パロディや二次創作としてオリジナリティを発揮する場合等の例外的な著作物の使用が認められるケースに該当すると考え掲載しています。 3. 当該学習データセットは、わいせつな視覚コンテンツを直ちに表示されないよう配慮されており、学習データセットによる投稿が公共の場においてわいせつな内容を表示することを意図するものではありません。 なお、著作権者の意に反する利用や不当な利益を得る利用は避けるべきであり、著作権保有者によって正当な理由をもって当該データセットの関係する過半数の関係者から申し立てされた場合、掲載は速やかに中止されることが保証されます。 また本人は、使用する機械学習データセット中に、著作権者の権利を十分に尊重していないと評価され得る可能性を持つ情報源からのデータが含まれ得ることを認識しており、開発者等関係者が依拠すべき法律の整備等の動向について適切な監視と管理の義務を負っています。 本人は、法律や倫理に反する行為を意図しておらず、著作権者を含むイラストレーター及びアーティスト等の権利を尊重し、人工知能の開発を進めることに努めます。</details>
|