CivArchive
    Krea2 StoryFlow — 3-Frame In-Context LoRA - v0.1
    Preview 140911914
    Preview 140911927
    Preview 140911968
    Preview 140911974

    原理

    参考FLUX In-Context LoRA训练方案训练的Krea2 三帧故事板,仍属于实验中 。

    把三个连续状态放在同一张纵向画布中联合生成,让模型利用画布内上下文保持人物、服装、场景和道具连续,同时表现动作、位置或镜头变化。

    使用方法

    推荐参数:

    Base: Krea2 Turbo BF16
    
    LoRA strength: 0.85(可尝试 0.7–1.0)
    
    Resolution: 512×1536
    
    Steps: 8
    
    CFG: 1.0
    
    Sampler: Euler Ancestral
    
    Scheduler: Beta

    提示词格式:

    [SHOT-SEQUENCE] 三帧共同保持的人物、服装、场景、道具和光线。
    
    [FRAME-1] 初始状态。
    
    [FRAME-2] 中间状态。
    
    [FRAME-3] 最终状态。

    建议先在 512×1536 生成(训练分辨率),再把三格分别超分。不建议直接生成超高纵向分辨率。

    当前问题

    • 小幅凝视、表情和手势变化有时不明显。

    • 复杂手部与道具交互可能出错或重复道具。

    • 强度过高时动作会变保守,后两格可能过于相似。

    • 远景人物的脸部细节有限。

    Description

    FAQ

    Comments (5)

    olivettyAug 26, 2026· 3 reactions
    CivitAI

    Really cool! Wonder if this would help character sheets!

    Coffmeteor
    Author
    Aug 26, 2026

    This is not specially made for character cards, but of course, I have specially trained to make specific character cards, used to make videos, but I did not upload them, because it is not very general and not very mature, lol

    jujucom5000374Aug 30, 2026
    CivitAI

    Great job, but how to use it?

    Coffmeteor
    Author
    Aug 31, 2026

    I've been a little busy lately, wait for meI've been a little busy with work recently. I'll improve the usage examples when I'm free, and other matching LoRa

    wyxzddsjj919Sep 2, 2026

    @Coffmeteor 佬,是生成后导入H3,用某些特别图片加载点(可以划分区域的)把它分别作为 首 中 尾 3张图吗

    LORA
    Krea 2

    Details

    Downloads
    295
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/26/2026
    Updated
    9/2/2026
    Deleted
    -

    Files

    krea2-meridian-storyboard-v003-final.safetensors