This lora depicts the character Takamiya Rion(鷹宮リオン) from Nijisanji

Trigger Word:Takamiya Rion, anime-style girl. The image features an anime-style character with long blonde hair styled in twin tails, each adorned with a red bow. The character has striking purple eyes and is wearing a white blouse with black trim and gold buttons, paired with a black skirt that has a frilled hem,

Trained: musubi-tuner, 30 images

Caption: adjusted caption from Qwen2.5-VL-3B (includes described background)

lr:(2e-4,flow_shift7.0)

epoch:100

train_step:3000

lora_strength:0.8

sample prompt:Takamiya Rion, anime-style girl. The image features an anime-style character with long blonde hair styled in twin tails, each adorned with a red bow. The character has striking purple eyes and is wearing a white blouse with black trim and gold buttons, paired with a black skirt that has a frilled hem, (she is dancing:2.0), (chibi Nendoroid:3.0) in bedroom.

musubi-tuner blocks_to_swap train speed:

4070Ti(VRAM12GB), RAM16GB, 544x544 resolution image training

no block_swap : 250s/it (watch youtube)

no block_swap : 190s/it (no watch youtube)

--blocks_to_swap 2 : 164s/it (no watch youtube)

--blocks_to_swap 6 : 126s/it (no watch youtube)

--blocks_to_swap 10 : 15s/it (no watch youtube)

--blocks_to_swap 12 : 3s/it (no watch youtube)

--blocks_to_swap 14 : 3s/it (watch youtube)

自分はこのblocks_to_swap設定のおかげでmusubi tunerの学習率を1/5に下げて実行することができるようになりました。

もっとも前回の設定(1e-3,sigmoid)だと300stepで済みましたが、この設定(2e-4,flow_shift7.0)だと300stepは全然足りず、1500stepでもちょっと足らず、3000stepが今回の結果になります。

このため前回の設定の学習stepの10倍くらいに設定しないといけないかもしれません。それでさえもLoRA強度は前回よりも弱く、その分動きが反映されやすい。

I can set lower learning rate(1e-3,sigmoid => 2e-4,flow_shift7.0) of musubi tuner due to this blocks_to_swap configuration.

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.