#school_uniform | Explore Tumblr Posts and Blogs

centrally-unplanned · 2 years

Text

I have a particular affection for the ways non-cultural designs & choices can create cultural artifacts. Anyone remember ToHeart2?

An ero visual novel coming in at the tail end of the VN era Tokimeki Memorial ushered in 1994 and Kanon crystalized in 1999, it was released in December of 2004 and was the ‘it’ game of its genre for the next year. It got an anime release in 2005, spin-off games, OVA’s, drama CD’s, the works. It was really big at Comiket in the next year, selling ‘image CDs’ of risque art, promotional manga, and a…car billboard?

ToHeart2 did not last though. Its known among visual novel fans for sure, but it's currently ranked #609th on Visual Novel Database in popularity - totally respectable, but no one is recommending it today as the “oh you gotta play this” game and if you aren’t into visual novels it is not coming up in your day-to-day anime-fandom conversations the way Clannad or Doki Doki Literature Club is. It had its era and the fandom moved on, like almost any other property.

But I bet way more people have encountered it than they know, because something else was released pretty much right after ToHeart2 was, in May of 2005 to be precise:

Oh yeah, this is that kind of essay. Danbooru is the original ‘successor’ to the 2ch-style imageboards that was art-first, non-hierarchical, and easily searchable, making it absolutely perfect for porn. The style quickly swept the competition and for a while Danbooru, and its architecture-clones like Gelbooru, ruled the hentai roost on the internet. If it's the year 2010 and you want some pics of Asuka in a school uniform stepping on you, Danbooru was your very first stop.

How you found those pictures is important here; Danbooru is 99% of the time used via searching for tags. Those who don’t have an account - aka most users - can only search by two tags, so you do the best you can; to take the above prompt you go “soryuu_asuka_langley stepped_on” and bam:

Only 4 results, are you kidding me?? Jesus get it together ero artists, some low hanging fruit here. Let's go with something more popular:

There we go, “red_hair school_uniform”, that is more like it. Okay, so sometimes you want Sailor Pluto + Zettai Ryuoiki or bust and you are gonna get what you are gonna get with those search terms, but other times you are exploring, you wanna see what you will find, any red-head will do, and you get this absolute sea of results. What are your options for navigating it?

Pretty…basic, right? And you can’t easily narrow down your search, you only get two search terms. So you do what comes natural, you look at each page, select out images you like, and hit next. But that is…a lot of pages, you aren’t going to read them all.

A funny thing about UI design is that if a button exists, humans are going to press it. We are just inherently curious, we want to try things. And this UI essentially gives us three buttons: “First”, “Next”, and ….“Last”, in this case Page 43654. So as you are building your stache of auburn-topped white button-ups, you are gonna click Page 1, then Page 2, then Page 3…and at some point your curiosity will get the better of you, and you are gonna click Last

And the posts are organized chronologically by order of submission. You get it now, right?

For large searches, which the two-tag cap encourages users to do, Danbooru’s UI gives you the option of showing you the *latest* posts…and the *first posts ever submitted to Danbooru*. If its 2013 you are not gonna stumble on anything from 2008 without actively looking for it, but there is always a button flashing in your face to show you every post from 2005.

When ToHeart2 was the most popular ero game around.

As you can bet Danbooru on opening day was flooded with ToHeart2 images. It generated so much content applicable to this site concept too - it's a split path harem game! Look at these girls:

You got your big titty red-head, white-haired soft kuudere, your lolis your genki’s, this is the OVA so it isn’t even all of them; option coverage is the point of these games. And it clocks in at, help me out here VNDB…

55 hours are you fucking kidding me?! This is the Anna Karenina of visual novels, this work is someone’s magnum opus, my god. Point is it had a lot of art - and you add in the anime, the Comiket CD’s, the promotionals, the manga, the doujin, its a lot. Wait, I don’t have to guess:

Its 6.2k images worth of ecchi art as of today. Every single ‘basic category’ search term people would commonly use had some ToHeart2 character in it, stored at the back of the history log. Far beyond its popularity as a visual novel, due to the inherent UI design of the site, anime fans got all the ToHeart2 porn they could handle just organically. And as time went on the ‘gap’ in between today and 2005 grew, cementing this little cultural artifact in place. Obviously other works from 2005 got similarly cemented, ToHeart2 is just the most prominent one to me.

I love how some of them have their context just buried in there too, like this one I pulled from the site:

This is the cover of a Comiket 70 doujin collection - actually a mix of characters, not ToHeart2-specific - by doujin group Spread-Pink, a talented group; something the median non-Japanese searcher saving this to their hard drive would have no clue about. Here is a link to the full doujin, it has Haruhi SOS Brigade propaganda, absolutely choice. This is of course a trait inherent to a lot of the older art on the internet - so much of it was taken, often without attribution or with edits, from random professional-adjacent works and repurposed into the context the internet needed it to be. Today the internet of Japan & the West is way more merged, and tracing back an image is way easier, but in the early days it was anarchy, and so early Danbooru is full of that, in comparison to the modern uploads.

And yes I cropped the image here, but it was for sizing purposes I swear, don’t judge me, the full one is in the link!

I do have one more reveal before I wrap this up, however - I have been lying to you. I doctored that Danbooru screenshot up there, because that isn’t how the search in Danbooru looks today. It looks like this:

Yup, the Last Page button is gone. I pulled a screenshot from the Wayback Machine for the edit, I am not sure when it vanished but I think it is sometime around 2015? Now this phenomenon lasted longer than that - at that time Danbooru clone Gelbooru had grown to higher popularity due to ease of use but imported all of Danbooru’s content including its chronology meta-data when it was made in 2007, and it kept the button. Gelbooru grew to such sizes that in the late 2010’s however, if you try to hit Last Page on a large enough search, you get:

Alas, a pagination cut-off (unless you upgrade your account, which is like paying for WinRAR).

So this isn’t a present cultural artifact, but a past one - a generation of internet users for a decade+ had privileged access to the ToHeart2 Hentai Collection, but in the 2020’s it's gone now, lowered down to equality with every other historical work. Alas ToHeart2, your reign was brief but beautiful, and may we all hope for an elysian future where Danbooru re-implements the Last Page button and your rule is restored.

#anime essay #history of the internet #meta-hentai #toheart2

55 notes · View notes

ebycow · 1 year

Text

LoRA学習方法簡易解説

1.ファイルとフォルダ構成について

ファイルツリーを以下に示す

images ├── regular │ └── 120_girl │ ├── transparent_10.png │ ├── transparent_11.png │ ├── transparent_12.png │ ├── transparent_13.png │ ├── transparent_14.png　... （正則化画像数>=教師画像数であれば多めに作成してよい）

├── tanaka │ └── train │ └── 120_tanaka girl │ ├── 001.caption │ ├── 001.jpg │ ├── 002.caption │ ├── 002.jpg │ ├── 003.caption │ ├── 003.jpg │ ├── 004.caption │ ├── 004.jpg │ ├── 005.caption

├── shimarin │ └── train │ └── 120_shimarin girl │ ├── 0.caption │ ├── 0.png │ ├── 10.caption │ ├── 10.png │ ├── 11.caption │ ├── 11.png │ ├── 12.caption │ ├── 12.png

1.1. まず，正則化画像フォルダと教師画像フォルダの二つが必要である． "教師用画像フォルダ"は "{繰り返し回数}{トリガーワード} {クラスワード}"　と設定し，このフォルダ名が「処理時に手続きの一部で使用される」 "正則化画像フォルダ"は "{繰り返し回数}{クラスワード}"と設定し，このフォルダ名が「処理時に手続きの一部で使用される」各{繰り返し回数}は「学習用画像の繰り返し回数×学習用画像の枚数≧正則化画像の繰り返し回数×正則化画像の枚数」となるように設定すること．

補記: ここでは便宜上フォルダ構成を/images/tanaka/train/ あるいは /images/shimarin/train としているが，trainフォルダ内に一つのトリガーワードを表すフォルダがあればよい(eg. /images/train/120_sks girl でもよい)

も��トリガーワード（一般的にはidentifierと呼ぶ）とクラスワードの関係を理解していない場合，関連する文献を調査し理解すること．

2.画像の用意

2.1.正則化画像を用意する一部のコミュニティでは，透明なpng画像を正則化に用いるのが最適であるといわれている．以下のリポジトリのプログラムは，Stable-diffusion-webUI上で透明画像を生成する拡張機能である，"正則化画像フォルダ"に透明画像を配置する． https://github.com/hunyaramoke/Generate-TransparentIMG

上述の通り，正則化画像数>=教師画像数であれば透明画像は多めに作成してよく, 余剰分は処理時に無視される

2.2. 教師画像を用意する教師画像を用意し"教師用画像フォルダ"に配置する．ファイル名は連番であるとよい．

3.キャプションの作成上記ファイルツリーの例では/images/tanaka/train/120_tanaka girl/001.caption のように各画像にキャプションを作成する．このcaptionファイルは学習時の調整に用いられる．captionファイルを用意しなかった場合，フォルダ名で指定した{トリガーワード} {クラスワード}がcaptionファイルの代替に指定される．（これは著しくモデルの精度を損なうことになる）

NAI形式のcaptionファイルの内容を以下に示す

1girl, breasts, looking_at_viewer, short_hair, bangs, multiple_girls, skirt, simple_background, shirt, hair_ornament, long_sleeves, white_background, ribbon, holding, 2girls, closed_mouth, school_uniform, blue_hair, jacket, yellow_eyes, white_shirt, pantyhose, pleated_skirt, collared_shirt, mole, blue_skirt, black_pantyhose, mole_under_eye, black_ribbon, neck_ribbon, letterboxed, instrument, music, guitar, shirt_tucked_in, playing_instrument, holding_instrument, electric_guitar, bass_guitar, gibson_les_paul

つまり，画像の要素をDanbooruタグ形式で列挙する，これを https://github.com/kohya-ss/sd-scripts で行う場合以下のようなコマンドとなる

python tag_images_by_wd14_tagger.py --batch_size <バッチサイズ> <教師データフォルダ>

python tag_images_by_wd14_tagger.py --batch_size 4 /images/tanaka/train/120_tanaka girl

手動で列挙出来る場合は，手動でもよい

4.学習学習は kohya-ss/sd-scripts https://github.com/kohya-ss/sd-scripts で行う．環境導入はこのプロジェクトがGradioでGUI化された kohya sd-scripts webui https://github.com/ddPn08/kohya-sd-scripts-webui で行い，インストールされたvenv環境を用いると手間がかからない．

学習時のコマンド例を示す．

../kohya-sd-scripts-webui/venv/bin/accelerate launch --num_cpu_threads_per_process 1 train_network.py --pretrained_model_name_or_path="JosephusCheung/ACertainty" --train_data_dir="/images/shimarin/train" --output_dir="/output/lora_shimarin" --reg_data_dir="/images/regular" --prior_loss_weight=1.0 --enable_bucket --min_bucket_reso=384 --max_bucket_reso=1280 --train_batch_size=12 --learning_rate=1e-4 --text_encoder_lr 5e-5 --use_8bit_adam --xformers --mixed_precision=fp16 --save_every_n_epochs=1 --save_model_as=safetensors --clip_skip=2 --seed=42 --color_aug --flip_aug --network_module=networks.lora --resolution=768,512 --network_dim 256 --max_train_epochs 10 --shuffle_caption

上記コマンドの補足： train_data_dir, output_dir, reg_data_dirに注意する．

pretrained_model_name_or_pathは"JosephusCheung/ACertainty"を用いるのが最適であるが，どれでもよい

network_dimは，コミュニティでは256が用いられるが，それ以下のほうが高速に学習できる．

--color_aug はモデルのVAEを用いて色相によりデータ拡張する

--flip_aug は画像を反転してデータ拡張する

--enable_bucket --min_bucket_reso=384 --max_bucket_reso=1280 はNovelAIのaspect-ratio-bucketing https://github.com/NovelAI/novelai-aspect-ratio-bucketing　の実装であり，学習時のアスペクト比を可変にする事が可能となるものである．つまり，学習時に画像の解像度をリサイズする加工は必要ない．

以下にヒントを引用する https://github.com/kohya-ss/sd-scripts/blob/main/fine_tune_README_ja.md

機械学習では入力サイズをすべて統一するのが一般的ですが、特に制約があるわけではなく、実際は同一のバッチ内で統一されていれば大丈夫です。NovelAIの言うbucketingは、あらかじめ教師データを、アスペクト比に応じた学習解像度ごとに分類しておくことを指しているようです。そしてバッチを各bucket内の画像で作成することで、バッチの画像サイズを統一します。

オプションをメモとして示す．

usage: train_network.py [-h] [--v2] [--v_parameterization] [--pretrained_model_name_or_path PRETRAINED_MODEL_NAME_OR_PATH] [--train_data_dir TRAIN_DATA_DIR] [--shuffle_caption] [--caption_extension CAPTION_EXTENSION] [--caption_extention CAPTION_EXTENTION] [--keep_tokens KEEP_TOKENS] [--color_aug] [--flip_aug] [--face_crop_aug_range FACE_CROP_AUG_RANGE] [--random_crop] [--debug_dataset] [--resolution RESOLUTION] [--cache_latents] [--enable_bucket] [--min_bucket_reso MIN_BUCKET_RESO] [--max_bucket_reso MAX_BUCKET_RESO] [--bucket_reso_steps BUCKET_RESO_STEPS] [--bucket_no_upscale] [--reg_data_dir REG_DATA_DIR] [--in_json IN_JSON] [--dataset_repeats DATASET_REPEATS] [--output_dir OUTPUT_DIR] [--output_name OUTPUT_NAME] [--save_precision {None,float,fp16,bf16}] [--save_every_n_epochs SAVE_EVERY_N_EPOCHS] [--save_n_epoch_ratio SAVE_N_EPOCH_RATIO] [--save_last_n_epochs SAVE_LAST_N_EPOCHS] [--save_last_n_epochs_state SAVE_LAST_N_EPOCHS_STATE] [--save_state] [--resume RESUME] [--train_batch_size TRAIN_BATCH_SIZE] [--max_token_length {None,150,225}] [--use_8bit_adam] [--mem_eff_attn] [--xformers] [--vae VAE] [--learning_rate LEARNING_RATE] [--max_train_steps MAX_TRAIN_STEPS] [--max_train_epochs MAX_TRAIN_EPOCHS] [--max_data_loader_n_workers MAX_DATA_LOADER_N_WORKERS] [--persistent_data_loader_workers] [--seed SEED] [--gradient_checkpointing] [--gradient_accumulation_steps GRADIENT_ACCUMULATION_STEPS] [--mixed_precision {no,fp16,bf16}] [--full_fp16] [--clip_skip CLIP_SKIP] [--logging_dir LOGGING_DIR] [--log_prefix LOG_PREFIX] [--lr_scheduler LR_SCHEDULER] [--lr_warmup_steps LR_WARMUP_STEPS] [--prior_loss_weight PRIOR_LOSS_WEIGHT] [--no_metadata] [--save_model_as {None,ckpt,pt,safetensors}] [--unet_lr UNET_LR] [--text_encoder_lr TEXT_ENCODER_LR] [--lr_scheduler_num_cycles LR_SCHEDULER_NUM_CYCLES] [--lr_scheduler_power LR_SCHEDULER_POWER] [--network_weights NETWORK_WEIGHTS] [--network_module NETWORK_MODULE] [--network_dim NETWORK_DIM] [--network_alpha NETWORK_ALPHA] [--network_args [NETWORK_ARGS [NETWORK_ARGS ...]]] [--network_train_unet_only] [--network_train_text_encoder_only] [--training_comment TRAINING_COMMENT]

詳細は，https://github.com/kohya-ss/sd-scripts/blob/main/README-ja.mdを熟読すること．

5.トラブルシューティング python3.8で学習時，train_network.pyにて TypeError: 'type' object is not subscriptable というエラーが表示される https://github.com/kohya-ss/sd-scripts/issues/38 に記載がある．

The error seems to be related to type hints feature of Python. ->list[LoRAModule]: This notation seems to be supported from Python 3.9. This repo is tested on Python 3.10, but you seem to use Python 3.8. Please update Python to 3.10.x. If it is difficult, please remove ->list[LoRAModule]. However, another error might occur, so I recommend to use Python 3.10.

#stable-diffusion #LoRA

1 note · View note