LORA
Wan Video 2.2 T2V-A14B
LORA
Wan Video 2.2 T2V-A14B
Wan2.2 T2V 14B

Shiina Yuika/椎名唯華 | Wan2.1 & 2.2 14B

kihh
Creator
⭐ 0.0
⬇ 186 Downloads
👁 1 Views
🖼 15 Images

Trained Words

Shiina Yuika
anime-style girl. The image features an anime-style character with pink hair styled in a short bob cut. The hair is slightly tousled and has a few strands falling over the shoulders. The character's eyes are large and expressive
with a light blue color that complements the overall pastel aesthetic of the image.The character is wearing a white sweater with a red bow tie around the neck. She is dressed in a school uniform consisting of a white blouse with a red bow tie
a beige cardigan
and a red and white plaid skirt.

About this model

-------------------------

Wan2.2:

Trigger Word:Shiina Yuika, anime-style girl. The image features an anime-style character with pink hair styled in a short bob cut. The hair is slightly tousled and has a few strands falling over the shoulders. The character's eyes are large and expressive, with a light blue color that complements the overall pastel aesthetic of the image.The character is wearing a white sweater with a red bow tie around the neck. She is dressed in a school uniform consisting of a white blouse with a red bow tie, a beige cardigan, and a red and white plaid skirt.

lr:(3e-4,flow_shift7.0,timestep_boundary 885)

epoch:140

train_step:4200

lora_strength:0.8

sample prompt:Shiina Yuika, anime-style girl. The image features an anime-style character with pink hair styled in a short bob cut. The hair is slightly tousled and has a few strands falling over the shoulders. The character's eyes are large and expressive, with a light blue color that complements the overall pastel aesthetic of the image.The character is wearing a white sweater with a red bow tie around the neck. She is dressed in a school uniform consisting of a white blouse with a red bow tie, a beige cardigan, and a red and white plaid skirt. (she is dancing:4.0), (chibi Nendoroid:4.5) in bedroom.

アクションは強調promptによって3.0~5.0に調整する必要があります。またはモーションloraを併用ください。

lightx2v Wan2.2 v1.1との併用を推奨します。

Action prompt should be adjusted to 3.0 to 5.0 depending on emphasis prompt. Or motion lora use together.

-------------------------

Wan2.1:

It is recommended to use with lightx2v Wan2.2 v1.1 lora.

This lora depicts the character Shiina Yuika(椎名唯華) from Nijisanji

Trigger Word:shiina yuika nijisanji, anime style girl. An anime-style character with pink hair and a casual school outfit, featuring a beige cardigan, red bow tie, and plaid skirt

Trained: musubi-tuner, 20 images

Caption: adjusted short caption from ChatGPT (includes described background)

lr:(2e-4,flow_shift3.0)

epoch:230

train_step:4600

lora_strength:0.9

sample prompt:shiina yuika nijisanji, anime style girl. An anime-style character with pink hair and a casual school outfit, featuring a beige cardigan, red bow tie, and plaid skirt, (she is dancing:1.8), (chibi Nendoroid:2.5) in bedroom.

アクションは強調promptによって2.0~4.0に調整する必要があります。またはモーションloraを併用ください。

Action prompt should be adjusted to 2.0 to 4.0 depending on emphasis prompt. Or motion lora use together.

以下備考:

・キャラクター再現度とpromptの動きやすさはトレードオフで、おそらく生成プロンプト(トリガーワード)が長いほど再限度は上がるが、動きやすさは制限される。

動きにくい場合はlora強度を下げるか強調promptが有効だがKijaiのワークフロー( のワークフロー使用)だとprompt強調や{A|B|C}構文が上手く働かなかったのでTextEmbed Bridgeを使用するよう改造した。

とはいえ今回のキャプションは短い目なので強調promptを未使用でも比較的動きは反映されやすいです。

・musubi-tunerの学習でHunyuanVideoよりblocks_to_swapを大きくしないとVRAM不足に陥りやすそうだった。

しかし、Wan2.1のlora学習速度はHunyuanVideoと同程度だった。HunyuanVideoは14くらいで十分だったが、Wan2.1は24くらいにする必要があった。

4070Ti(VRAM 12G) musubi tuner

--blocks_to_swap 14 …198.27s/it

--blocks_to_swap 24 … 3.3s/it

・推移損失:

avr_lossは開始時0.100

100epoch(2000step)で0.085

230epoch(4600step)で0.065

250epoch(5000step)で0.060

・文字が書けるか

漢字は書けるようだ。ひらがな、カタカナは書けなかった。代わりに変な記号がでる。

あと漢字を書くと何故か画質が落ちました。

・生成時間はHunyuanVideoの2~3倍かかるのでLoRAがあるならHunyuanVideoのほうが生成ガチャ出来る分有利かも…。

HunyuanVideoで上手くいかなかったpromptをWanが描けるのかはまだ試せてないです。

Related Models

Similar AI models you may like