CivArchive
    YUNSANG Style [MOSS // Anima] - Anima v1.0-typeB
    NSFW
    Preview 143780902
    Preview 143780905
    Preview 143780900
    Preview 143780907
    Preview 143780910

    本页面包含 MOSS 和 Anima 两个版本,请选择对应的使用说明。
    This page includes both MOSS and Anima versions. Please follow the instructions for your model.

    _______________________________________________________________________________________

    FOR ANIMA:

    • 触发词/Trigger word: @yunsangstyle

    • 权重/Recommended weight:0.7~1 (V2 Recommended:0.6-0.8)

    • Sampler: er_sde + beta OR euler + normal

    • CFG:4-5

    FOR MOSS:

    base较好,但sft也尚可,base are better,but sft also ok.

    复杂场景,请提高CFG和步数,SFT例外。For complex scenes, increase the CFG and sampling steps,except when using SFT.

    • 触发词/Trigger word: @yunsangstyle

    • 权重/Recommended weight:1~1.2

    • Sampler: er_sde + beta OR euler + simple

    • CFG:3-5

    • STEPS:30-50

    ____________________________________________________________________________________

    MOSS

    V1

    • 和ANIMA同数据集训练,试了二十炉,不同配方,这一版是相对较好的,但感觉依然欠拟合,我大概会再作尝试。

    • Trained on the same dataset as ANIMA. I tried 20 runs with different training recipes, and this version is one of the better ones. It still feels underfit, though, so I’ll probably keep experimenting.

    • 请测试,如果有什么问题,请留言告知。Please test it and let me know in the comments if you encounter any issues.

    base复原能力最好,但sft在某种情况上更好,但基本上在多数情况base更还原画风。The Base model reproduces the style best overall. SFT performs better in certain cases, but Base is more faithful to the original art style in most situations.

    ____________________________________________________________________________________

    ANIMA

    v2:

    • target_res 1536,P-LORA ,修改了参数,指令遵从度上升,更好的泛化,光影提升、画面表现更好,增加少量数据集和局部参考。

    • Retrained at target_res 1536,P-LORA with adjusted training parameters, resulting in improved prompt adherence, better generalization, enhanced lighting and shadows, and higher overall image quality. A small amount of new training data and additional local reference images were also incorporated.

    _________________________________________________________________________________

    v1:

    • target_res 1536 重训,更好更细腻的纹理,和更好的泛化,皮肤质感提升极大。

    • Retrained with target_res 1536. Better and more refined textures, better generalization, and a huge improvement in skin texture.

    TYPE A vs. TYPE B

    typeA:更好的泛化,没有那么强烈的脸部和体型风格,配搭角色能够呈现出更多的脸型、体型、动作、构图,但相应皮肤质感没有B那么好。

    typeB:使用新画风3权重,老画风1权重的方式训练,强烈的风格,新画风更多的表现,如果泛化不好,请调低权重。

    如果你想要更强烈的风格,选B,如果想搭配更多的LORA使用,和更多样化,A会更好。

    TYPE A vs. TYPE B

    TYPE A — More flexible
    Better generalization with a less pronounced facial and body style. When paired with character LoRAs, it allows for a wider variety of face shapes, body types, poses, and compositions. The trade-off is that its skin texture is less refined than TYPE B’s.

    TYPE B — Stronger style
    Trained with a 3:1 weighting of the new art style to the old one, making the new style more prominent. If it doesn’t generalize well with your prompts, try lowering the LoRA weight.

    Which should you choose?
    Choose TYPE B for a stronger visual style. Choose TYPE A if you want to combine it with more LoRAs or create more varied results.

    Description

    A more distinctive style with finer skin texture and richer lighting. If it feels less flexible with your prompts, try lowering the LoRA weight.

    FAQ

    Comments (6)

    pozkerOct 7, 2026
    CivitAI

    我似乎遇到了胸部裸露度问题,提示词明确描述衣装的情况下,出图结果还是很容易偏向胸部大面积裸露,特别容易露出下半乳房(南半球)。

    table_oH
    Author
    Oct 7, 2026

    which one?

    pozkerOct 9, 2026

    @table_oH typeA

    table_oH
    Author
    Oct 9, 2026

    @pozker 你有没有尝试将下半球放到反面提示词里呢?或者说,是固定那套衣服才有问题,抑或是全部衣服都有问题呢?

    pozkerOct 9, 2026

    @table_oH 在我尝试将下半球放到反面提示词后,这点有生效,抱歉,是我脑子抽筋了,连这么简单的事情都没有提前想到,打扰你了。不过我的提示词为修女风格的服装,也尝试过多次修改细节描述,但是确实很容易导致最后生成的画面中角色胸部裸露度偏高,我正在逐一排查各种可能性

    table_oH
    Author
    Oct 9, 2026

    @pozker 因为本来就是裸露度高的风格,所以你可以多尝试反面提示词,素材集那边基本都是裸露度高的衣服。

    LORA
    Anima

    Details

    Downloads
    732
    Platform
    CivitAI
    Platform Status
    Available
    Created
    9/25/2026
    Updated
    11/20/2026
    Deleted
    -
    Trigger Words:
    @yunsangstyle

    Files

    yunsang_v1_typeb.safetensors

    Mirrors

    CivitAI (1 mirrors)