generated at
web UIでTextual Invertionする

>【】内はデフォルト値

0 画像を集める
データの質が正義
なので頑張って集めましょう
Trainタブへ移動

1 空のembeddingを作る
📌 Create embedding タブ
Name
これがトリガーワードになる
(生成するときはこの名前をプロンプトに書く)
Initialization text_初期化テキスト
最初はこのテキストがベクトルを埋め尽くす
トレーニングしなかった場合(今回の例だと)
プロンプトに quoria-san を入れても、 chibi を入れたのと同じになる
目的に近い言葉を入れたらよい?
Number of vectors per token
トークンをどれくらい使うかの指定
txt2imgでトークンは75個しか指定できない
あまり大きな数字をいれるとそれだけでトークンを使い切ってしまうので注意
じゃあ何個が最適なの?
👉 Create embedding

1.5 下準備(任意)
データセットのサイズを整えるなどの下準備をここですることができる
📌 Preprocess images タブ
Source directory
Destination directory
Width / Height
Existing Caption txt Action
その他
👉 Preprocess

2 トレーニング
📌 Train タブ
Embedding
1で作った空embeddingモデルを指定
Embedding Learning rate【0.005】
下げるほど複雑な学習ができるが、時間がかかる
必要であれば0.00005に下げてもいい
Gradient Clipping【disabled】
Batch size【1】
🚨Dataset directory
データセット画像が入ったフォルダを指定
Log directory
Prompt template
Width / Height【512/512】
Max steps【100000】
中断/再開ができるので大きくしとけばいい(らしい)
Save an image to log directory every N steps, 0 to disable【500】
N回に一回トレーニングしたembeddingsで画像を生成する
うまくいっているかの確認
Save a copy of embedding to log directory every N steps, 0 to disable【500】
上と大体同じ
Save images with embedding in PNG chunks【✅】
Read parameters (prompt, etc...) from txt2img tab when making previews【🟩】
Shuffle tags by ',' when creating prompts.【🟩】
Drop out tags when creating prompts.【0】
Choose latent sampling method【once】
👉 Train Embedding

3 使ってみる
txt2imgでプロンプトにモデル名を入れて生成する
TI-quoria-san embeddings
/qualia-sanから30枚選んでTI
Number of vectors per tokenは15で他はほぼデフォルト
ACertainModelで試す
$ 1girl,TI-quoria-san,(wearing yukata:1.2)
$ 1girl,TI-quoria-san,(red hair:1.4)
トリガーワードは強いので、弱めたほうがいい?
$ 1girl,(TI-quoria-san:0.7),red hair
$ 1girl,TI-quoria-san,(wearing yukata:1.2)
普通に8時間くらいかかったnomadoor
多分今回の場合は100000回も回さなくて良い
500stepsですでにクオリアさんになっていた
embeddingsってなんのモデルと組み合わせてもいいの知りませんでした…
めちゃめちゃ綴りまちっがてるな…(今更)