Type | |
Stats | 239 67 |
Reviews | (31) |
Published | Jun 14, 2024 |
Base Model | |
Training | Epochs: 80 |
Usage Tips | Strength: 1 |
Trigger Words | shiromuku kimono obi |
Hash | AutoV2 3AAF881CF0 |
6月の季節物
今回は日本伝統の花嫁衣装の白無垢
角隠しと綿帽子の区別はしてなく、例によってv001は顔が潰れやすい
v001 スクリプトのテスト
アップデートしてsqlデータベースから非推奨タグを推奨タグに置き換える処理を追加
その後dataset tag edditerで重複タグの削除とトリガー「shiromuku」とだけ追加して学習
画像から余計な情報のクリーンナップとかもやってない
v002
既存タグに対してクリーンナップしたタグを追加してたので修正したスクリプトを使用
v003
タグが大文字・小文字混在なのを全て小文字に統一
asian、 Chinese、 korean など文化圏を判別できてないタグの手動削除
sd-script/dev の '--alpha_mask' を使用
トリガーを 'shiromuku, 白無垢'へ、'japanese clothing'を'kimono'へ変更
v004
手動でタグ整理、とても手間な割には出来が満足とも言えない
角隠しを理解させようとその画像を増やしたが学ばない
ENG
Seasonal for June
This time, it's the traditional Japanese bridal attire, Shiromuku.
There is no distinction between tsunokakushi and wataboshi, and as usual, v001 tends to distort the face.
v001 Script Test
Image preprocessing script
Updated to include a process that replaces deprecated tags with recommended tags from the SQL database.
After that, duplicate tags were removed using the dataset tag editor, and only the trigger "shiromuku" was added for training.
No cleanup of unnecessary information from the images was done either.
v002
Added cleaned-up tags to existing tags, using a corrected script.
v003
Unified tags to lowercase as they were mixed case.
Manually removed tags like "asian", "Chinese", "korean" which could not distinguish cultural regions.
Used --alpha_mask
from sd-script/dev
.
Changed trigger from "shiromuku, 白無垢" to "shiromuku" and "japanese clothing" to "kimono".
v004
Manually organized tags; despite the effort, the results are unsatisfactory.
Increased images to make the model understand "角隠し" (horn concealing headdress), but it did not learn.