CivArchive
    PAIR OF VERY NORMAL PATENT LEATHER THIGH BOOTS|一双很普通的漆皮大腿靴 - v1.0-EP01
    NSFW
    Preview 31481917
    Preview 31481066
    Preview 31481102
    Preview 31481137
    Preview 31481213
    Preview 31481808
    Preview 31481832
    Preview 31481871
    Preview 31481961
    Preview 31482043
    Preview 31482155
    Preview 31482642
    Preview 31482663
    Preview 31791612
    Preview 31793930
    Preview 31794032
    Preview 31794710

    Yo~~~Long Time No See!! Guys!!

    It's Dr. LaoWai!! Da Real Fetisher Of BOOTS!

    Today I bring you a pair of very normal boots.

    YEAH,

    It's THIGH HIGH ! It's STILETTO ! It's POINTY TOE !

    Annnnnnd, it's super f**king shiny PATENT LEATHER !

    ==============================================================

    Are you annoying about when you enter the prompt 'THIGH BOOTS',

    The great_great_PONY_Checkpoint will always give you something you never asked for ?

    Like fold-over boots, lace-up boots, belt boots or just a V cut on the end of boots shaft that makes your boots really Fantsy/Ancient type?

    Or maybe you prompt 'THIGH BOOTS' , it just shows up a Calf-Length Boots ?

    Now, with this lora, I can 90% Ensure you, these problems are all gone !!

    Feel Free to use it with most pony based models, anime type and realistic type.

    · You don't need to lower the weight to 0.6 or 0.4 or 0.1, 0.8~1 should be great ! sometimes you even need to raise it more more more. Just try.

    · The Trigger Word is 'OBLK', cause I trained this lora based on the pictures of ITEM Code Name OBL-K-2355 from whatever online boots shop. Yet, Some Prompts Help. Like: thigh boots, boots, high heels, high heel boots, stiletto heels, black footwear, shiny footwear,

    ==============================================================

    This Lora is my third time trying use civitai lora training function.

    I set 19 images for dataset, and 20 epoch 10 repeats for training.

    It's quiet different from SD1.5 Lora training, so it's overcooked.

    only 000001 to 000005 epoch results are good.

    I test them on different pony based mix checkpoint, but it's hard to tell which epoch suit all checkpoints best. But 000001~000003 Epoch seems fit most anime pony based checkpoints.

    I also tried Lora Block Weight, again,it's way way different from SD1.5 Lora,and I just couldn't find any LBW sets recommand on Internet. So I started an experiment,to manually transform my SD1.5 LBW sets into SDXL LBW sets.Yet, it's not very stable,especially when you using multiple loras. So be aware.

    Description

    Less Infulence with other loras.

    FAQ

    Comments (3)

    navy123Oct 14, 2024
    CivitAI

    对靴子版型效果显著,不过靴筒部分还是有些随机。大佬这种服装部件的素材要什么样的,只是腿部还是也要全身?想炼一个短尖头的款。

    XQAnima
    Author
    Oct 14, 2024

    要掺杂着来,然后看你训练集图片总量。比如你搞了30张图片,其中20张是全身,4张是无人的商品摆拍特写,那不仅是无妨的,而且是极好的。有助于学习特征。而又比如C站上一位同行的绑带长靴LORA,他搞了54张训练集,但不是静态商品图,就是下半身的模特图,毕竟国外不像国内有XHS那么丰富的素材收集地。结果就是权重高于0.8他的lora就会只出下半身的图或者静物图。训练集数量越大影响越稳固,与其他泛化元素的拟合也会更好,比例调配的影响也会小。训练集数量少的话,特征训练会很快,但是容易过拟,也容易糊,训练集里面的分配比例,相差一两张图影响也会大。主要就这样。

    还有训练集不一定要完全不同到天差地别,你单图裁局部细节,其实也可以构成一个新的训练样本,用以模拟不同的构图情况,而完全不用担心过拟合。你还可以左右翻转,你还可以把原图导入PS画两笔,比如你手头只有商品的静物商品图,根本没有上脚的买家秀,那也无问题,打开PS,抠出长靴,粘到人体模特上,只要肢体错误不是特别严重,lora在学习解构后,重新生成的时候是会利用大模型里的其他元素辅助,帮助消除PS原图的不自然感的,这招也是我从老外那里学来的,手搓训练集。

    navy123Oct 15, 2024

    @XQAnima 哦哦哦,谢谢大佬指教,我去试试

    LORA
    Pony

    Details

    Downloads
    461
    Platform
    CivitAI
    Platform Status
    Available
    Created
    9/26/2024
    Updated
    6/12/2026
    Deleted
    -
    Trigger Words:
    oblk
    thigh boots
    black footwear
    stiletto heels
    shiny footwear

    Files

    OBLK-000001.safetensors

    Mirrors

    CivitAI (1 mirrors)

    Available On (1 platform)

    Same model published on other platforms. May have additional downloads or version variants.