CivArchive
    p1ct0a1- ckxl/zimage base - zimage base 1.0
    Preview 119105464
    Preview 119102644
    Preview 119102652
    Preview 119102649
    Preview 119102651
    Preview 119102647
    Preview 119102650
    Preview 119102643
    Preview 119102645

    3.0:触发词: photobashing
    ——————————————————————————
    2.0:触发词(trigger word):p1ct0a1,

    基于 zimage base (Based on the zimage base model)

    已知15000 步仍未过拟合,loss 收敛缓慢。 (Even after 15,000 training steps, the model has not yet overfitted, and the loss is converging slowly.)

    在 Turbo 上使用建议的 1–1.5 权重,效果只能说是勉强可用 (On Turbo, using the recommended weight of 1–1.5 is barely acceptable)
    ——————————————————————————
    1.0:这是我第一次尝试训练 LoRA,还在探索阶段,技术力有限。

    这个模型主要用于增强画面的照片合成质感与冷色调氛围。它能给原本平淡的画面增加一种独特的颗粒感、色散效果和电影级的叙事光影。

    推荐设置:

    触发词: photobashing

    推荐权重: 0.6 ~ 0.8

    采样器选择 (Sampler):

    追求意境与朦胧氛围:推荐使用 Euler 或 Euler a。

    追求清晰度与线条细节:推荐使用 DPM++ 2M。

    可选风格提示词 (Optional Tags):

    为了获得最佳效果,建议配合以下提示词使用:

    cyan theme, chromatic aberration, cool colors, blurry background, high contrast, noise, grunge

    如果是复古/黑白风格,可尝试:sepia, monochrome

    Introduction:

    This is my first attempt at training a LoRA. I am still learning the ropes, so any feedback is welcome.

    This model is designed to enhance the Photobashing texture and Cold/Cyan atmosphere of your generations. It adds a distinct layer of grain, chromatic aberration, and a cinematic storytelling vibe to otherwise flat images.

    Recommended Settings:

    Trigger Word: photobashing

    LoRA Weight: 0.6 ~ 0.8

    Sampler Recommendation:

    For Atmosphere & Vibe: Use Euler or Euler a.

    For Details & Sharpness: Use DPM++ 2M .

    Optional Style Tags:

    For the best results, try mixing in these prompts:

    cyan theme, chromatic aberration, cool colors, blurry background, high contrast, noise, grunge

    For vintage/B&W styles: sepia, monochrome

    Description

    FAQ

    Comments (10)

    RehvkaJan 29, 2026· 1 reaction
    CivitAI

    GOOD!!

    TANLI2023Jan 29, 2026
    CivitAI

    What is the value of Learning Rate?

    FEEEEF
    Author
    Jan 29, 2026· 1 reaction

    总学习率:0.0001


    conkikhon406Feb 1, 2026

    @FEEEEF too low

    ACatStickJan 29, 2026
    CivitAI

    大佬参数怎么写的

    FEEEEF
    Author
    Jan 29, 2026

    没改,用的默认
    保存设置

    每N轮保存一个LoRA:1

    LoRA保存精度:bf16

    学习率&优化器

    总学习率:0.0001

    优化器:AdamW

    数据集设置

    图片分辨率:1536 * 1536

    网络设置

    网络大小:32

    特色功能

    是否启用LoRA+:是

    LoRA+学习率倍数:3

    ACatStickJan 29, 2026

    @FEEEEF 唔这是在线训练?aitoolkit里没看见这种功能

    FEEEEF
    Author
    Jan 30, 2026

    @eubanksjhw209 在魔塔上面免费训练的

    GarlandFeb 3, 2026
    CivitAI

    如果你的素材集大概是几十张的话其实可能已经拟合了,我发现稍微改动提示词就会让风格大变(在特定的文本标注上已经拟合),这大概是文本标注的处理手段太单一导致的。

    FEEEEF
    Author
    Feb 3, 2026

    100张的训练集。
    如果拟合度不够高,在turbo的图片会变成真人画风的。
    后面25000步 0.001学习率的好了点,但loss依然稳定0.34左右。
    我是觉得这个版本是不够拟合的

    LORA
    ZImageBase

    Details

    Downloads
    405
    Platform
    CivitAI
    Platform Status
    Available
    Created
    1/28/2026
    Updated
    5/27/2026
    Deleted
    -

    Files

    p1ct0a1.safetensors

    Mirrors

    HuggingFace (1 mirrors)
    CivitAI (1 mirrors)