CivArchive
    Anime style lora M1 | Wan2.1 & 2.2 T2V 14B - Wan2.1 T2V 14B
    NSFW
    Preview 84904003
    Preview 84904040
    Preview 84910756
    Preview 84904129

    Trigger Word:anime-style girl.

    Trained: musubi-tuner, 20~30 AI illustration images

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

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

    epoch:100

    train_step:3000

    lora_strength:0.8

    sample prompt:Hatsune Miku, anime-style girl. (she is playing bass guitar:1.2) in bedroom.

    アクションは強調promptによって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 depending on emphasis prompt. Or motion lora use together.

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

    ・kijaiのCausvid v2_loraを使ってstep=20,cfg=6.0からstep=10,cfg=1.0としたとき

    動画の生成速度が4倍に上がったが生成品質が良くなかったり、promptのモーションが反映されにくくなった。

    promptのモーションが反映されにくい問題はprompt強調の倍率を上げるとpromptに従って動くようになった。

    キャラクタLoRAでもpromptが反映されにくいことがあったので同じ問題だろう。

    一方で生成品質があまりよくないのに対して試しにスタイルloraを作ってみることにした。

    一応、cfg>1を設定すると品質は上がるが、生成時間が倍に伸びる。cfg=1.0のまま質を上げたい。

    画像生成AIで作られた画像を集め、それを三つに分類してloraを三つ学習させた。

    M1はアニメ風、M2はシンプルなかわいいイラスト、M3は複雑な綺麗なイラスト。

    基本的にCausvid loraとの併用を推奨します。

    また複数loraを使用するためかloraの読み込みに一旦バグってしまった場合はComfyUIの再起動が必要でした。

    ・When using kijai's Causvid v2_lora with step=20,cfg=6.0 to step=10,cfg=1.0, the video generation speed was increased by a factor of 4, but the quality was not good and it was difficult to see the prompt motion.

    The problem of the prompt motion not showing up well was solved by increasing the magnification of the prompt emphasis, which made the motion follow the prompt.

    The same problem may have occurred with the character LoRA, since it was sometimes difficult to see the prompt motion.

    On the other hand, the quality of the generated images was not good, so I decided to create a style LoRA to try it out.

    I wanted to keep cfg=1.0 to improve the quality.

    I collected images created by the image generation AI and classified them into three categories to train three lora.

    M1 is an animation style, M2 is a simple cute illustration, and M3 is a complex beautiful illustration.

    Basically, it is recommended to use with Causvid lora.

    Description

    FAQ

    Comments (6)

    kim400Jul 17, 2025
    CivitAI

    hi! I tried your workflow but I can't find what wan-test18-000200.safetensors file is. What is this lora?

    kihh
    Author
    Jul 18, 2025

    This is character lora. This lora is https://civitai.com/models/1522560

    kim400Jul 18, 2025

    kihh oh thanks!!

    a1161327317Jul 23, 2025
    CivitAI

    How did the actions in the example picture come about? Is it another LoRA??

    a1161327317Jul 23, 2025
    CivitAI

    I don't quite understand the function of this LoRA. Does it make the results of image - to - video more like animations?

    kihh
    Author
    Jul 24, 2025

    If you are referring to the anime opening parody(Utena), this is a simple vid2vid, not i2v.

    I used Causvid_v2, character lora(I trained), style lora(This).

    This style lora only produces a Japanese anime painting style.

    LORA
    Wan Video 14B t2v
    by kihh

    Details

    Downloads
    284
    Platform
    CivitAI
    Platform Status
    Available
    Created
    6/27/2025
    Updated
    5/22/2026
    Deleted
    -
    Trigger Words:
    anime-style girl.

    Files

    Wan21_M1.safetensors

    Mirrors

    HuggingFace (1 mirrors)
    CivitAI (1 mirrors)