CivArchive
    ERNIE furry - ERNIE-furry-base_0504
    NSFW
    Preview 129714458
    Preview 129714464
    Preview 129714468
    Preview 129714483
    Preview 129714490
    Preview 129714505
    Preview 129714510
    Preview 129714521
    Preview 129714526
    Preview 129714533
    Preview 129714536
    Preview 129714547
    Preview 129714574
    Preview 129714589
    Preview 129714599
    Preview 129714612
    Preview 129714614
    Preview 129714648

    This model was trained on approximately 1100 images of various types, thus being suitable as a base knowledge, and may be subject to targeted training later.

    Data:The main part consists of diverse images used in previous fursuit models, which were mostly manually labeled.

    The second part is many images collected during the SD1.5 era, mostly automatically labeled, with a tendency towards thick coating style, and the prompts were not adequately checked.

    Then there are new collected images such as sdxl chroma, which have undergone sufficient checks. The sdxl images tend to be overly smooth. The chroma images are used for aesthetic training.

    Therefore, this LORA can be considered to be used to enable the model to acquire general knowledge, and is suitable to exist as a base model, allowing subsequent training models to have sufficient generalization ability.

    中文描述:

    模型在约1100张多种类型的图像上训练,因此适合作为基础知识存在,后续可能再进行定向训练。

    数据:

    主要是以前兽装模型所使用的多样化图像,这部分主要是手动打标。

    其次是许多在SD1.5时代收集的图像,多数自动打标,偏向厚涂风格,prompt没有足够的检查。

    然后是sdxl chroma 等新收集的图像,经过足够的检查,sdxl图像偏向过度平滑。chrome图像作为美学训练。

    所以此lora可以认为用于让模型获得通用的知识,适合作为一个底模存在,让后续训练模型有足够泛化能力。

    训练信息:

    训练分辨率为576x3->704x2->832x1,在训练时逐渐减少重要性不高的图像。这是出于训练速度和显存考虑。

    576训练所有图像。704减少不重要图像。

    832分辨率只训练了写实图像,最后只训练兽装图像。

    16G显存训练的分辨率上限大概是960。理论上1024会爆显存,因为训练太慢,因此不进行尝试。

    从turbo模型测试结果来看,只进行写实训练时对已经学到的动画能力不造成明显影响。多样性的美学训练动作训练效果不明显。

    训练程序:OneTrainer

    rank:64

    alpha:32

    在"Lora base model"中填入此模型时,

    这些参数能让你以此模型为底模,继续使用其它图像训练新的模型,而不必重新让模型学习基础概念。

    Dropout probability: 0.1

    resolution: 需要快速训练时建议默认的512,结构学习更容易。我使用的分辨率是受限于显存,而且实测高分辨率学习效果不佳,拟合缓慢。

    Timestep Shift: 默认1,使用3获得更快的学习速度,但图像变化更大,可能会破坏已有的学习成果。

    写实图像难以拟合,动漫风格则相对容易拟合。

    建议使用单一风格容易训练,多风格训练较为困难,可能是因为无法像sdxl那样训练文本编码器。

    Description

    first version, focused on fursuit.

    LORA
    Ernie
    by hxyy

    Details

    Downloads
    31
    Platform
    CivitAI
    Platform Status
    Available
    Created
    5/5/2026
    Updated
    5/14/2026
    Deleted
    -
    Trigger Words:
    kemono fursuit
    realistic fursuit
    toony fursuit

    Files

    ERNIE-furry-model_0504.safetensors

    Mirrors