CivArchive
    concept Sandwiched double penetration(locon) - v1.0
    NSFW
    Preview 462490
    Preview 462564
    Preview 462452

    首先我要承认这个lora不是那么容易使用,因为训练集确实比较难以搜集。生成一张令人满意的图片可能需要做不少尝试和筛选,目前最大的问题是经常会丢失脚部,但是我还是先把它发布出来再说。

    First of all, I have to admit that this lora is not so easy to use, because the training set is really difficult to collect. It may take a lot of trial and error to produce a satisfactory image, the biggest problem so far is that the feet are often missing, but I will post it first.

    由于一些原因,这个lora(locon)与我以往制作的不同。它并没有触发词,而是作为一个概念强化。你需要去使用prompt描述你想要的,如果你不知道怎么做,那么请看下列示例:

    This lora(locon) is different from the ones I have made in the past. It does not have a trigger word, but acts as a conceptual reinforcement. You need to use the prompt to describe what you want, if you don't know how to do it, then please see the following example:

    sandwiched, double penetration, group sex, 1girl, 2boys, mmf threesome, hetero, ass, veiny penis, blush, completely nude, vaginal, anal, leg up

    在此基础上你可以做出改动,比如删去"leg up"改为“standing,standing sex”(尽管这比较困难,也许你需要提高权重)、更换人物构成为“3girls,futanari,futa with female”(futa会让画面更不稳定,你可以使用mmf的图像的controlnet来辅助),不要忘了删除相反的prompt。

    On this basis, you can make changes, such as deleting "leg up" and changing it to "standing, standing sex" (although this is more difficult, maybe you need to increase the weight), changing the character composition to "3girls, futanari, futa with female "(futa will make the picture more unstable, you can use the controlnet of the mmf image to assist),don't forget to delete the opposite prompt.

    hiresfix是必要的,权重我一直使用1

    hiresfix is necessary, weight I always use 1.

    如果你发现使用人物lora后难以控制,我建议使用controlnet来进行辅助,首先生成一个不带人物lora的图片,然后将它放入controlnet中,最后加入人物lora后再次生成,以下为示例:

    If you find it difficult to control after using the character lora, I suggest using controlnet for assistance. First generate a picture without character lora, then put it into the controlnet, and finally add the character lora and generate it again. The following is an example:

    Description

    FAQ

    Comments (15)

    Saya
    Author
    Apr 10, 2023
    CivitAI

    刚刚误上传了错误的版本

    just uploaded the wrong file

    klaudiachyntia12581Apr 10, 2023· 9 reactions
    CivitAI

    sir.. can i request penis on armpit?

    Saya
    Author
    Apr 11, 2023

    need more better for another lora model sir please :(

    yagami_hayateApr 10, 2023· 2 reactions
    CivitAI

    danbooru搜索double penetration有8.3k之多,(当然,需要很多筛选的工夫),

    俺一直很好奇,danbooru早就有很多这样的tag,double penetration、grabbing another's hair、condom belt、tally,很多基本模型,waifu、anything,传说中他们是从danbooru学画画的,却不会画这些tag……

    ……其实俺想说,俺搜集了大概70张图片,是triple penetration的……如果过段时间可能会有上百张……

    俺不懂怎样从游戏拆cg包,俺只是对一些游戏截图,质量比较差,全是打码的,很模糊……俺计划过段时间就图生图来去掉码,俺还计划把棒棒做成x-ray那样的,目前都仅仅只是计划(新建文件夹)……

    如果楼主自己硬盘早就有一些游戏cg包,(当然cg包也是有码的),例如,

    guilty社的游戏,虏之xx系列,nightmare系列,

    画师m&m给atelier kaguya画的两个游戏,最终吃汉电车3、禁断病栋,

    画师m&m给astronauts sirius画的游戏,极限吃汉特异点123……俺还没开始截图这个系列,俺还没找到3的全cg档……

    俺还计划对custom maid 3d进行截图,可是他好像win10运行不了,大概俺还得折腾一阵子……

    ——总之,如果楼主有这些,或者楼主呆在哪个群的群友们有,那么楼主应该不需要俺的70个粗糙的截图……

    如果楼主大佬还是想要这些图,请告知俺怎样可以发给楼主……

    Saya
    Author
    Apr 11, 2023

    感谢支持,虽然double penetration图片总数很多,但是我希望制作的是这种前后包夹的视角,而danbooru上的double penetration一半以上是fellatio+vaginal,我需要的sandwiched比例非常少,筛选掉一些偏差过大的视角之后,最终使用的数据集不到100张,这其中还包括了一些futa、假屌。

    至于为什么这些tag没有被学习,我其实也一直不能理解......

    triple penetration的图片我还没有收集,如果您愿意,谢谢分享!

    yagami_hayateApr 11, 2023

    @Saya 邮箱发送啦~

    yagami_hayateApr 14, 2023

    @Saya 又有50图发邮箱啦……

    俺突然找到硬盘一些里番带了字幕组送的cg包赠品……不过很多片子只要看标题就知道是纯爱或者后宫的……

    请问,俺有一些黑白的漫画,而且漫画会写一些“扑哧扑哧”的拟声词,这是不是会污染训练?

    俺基本不会有更多的图啦……

    Saya
    Author
    Apr 14, 2023

    @yagami_hayate 非常感谢!拟声词必然会产生污染(有一些低质量的lora就很明显看出训练集处理的不好),但是如果比例少就不会有什么影响,我在制作lora的时候会尽量手动p掉图中的拟声词和对话框水印之类的杂质的(这个真的跟上班一样)

    ArturitoApr 12, 2023
    CivitAI

    Great work. How do you train these LoRAs? does it require a different method than training a character/style? Like a lower training rate and more training images?

    Saya
    Author
    Apr 12, 2023· 1 reaction

    I think concept(Action) lora need more steps and images than character lora.However, I have hardly trained the character lora. This point of view is based on the comparison with the character lora trained by other people.Because in the judgment of the character lora, it is more tolerant to the overfitting of the painting style.My training set is rarely less than one hundred pictures, and the step is more than two or three thousand.As for the learning rate, I did not make any special adjustments.

    ArturitoApr 12, 2023

    @Saya Thank you. I was also wondering if you used live-action (real) photos or drawings, and if you had to caption them extensively? And if you used Clip skip

    Saya
    Author
    Apr 12, 2023· 1 reaction

    @Arturito I rarely use real photos. Yes. Clip skip 2.

    emotionaldreams4May 17, 2023
    CivitAI

    would help i u mentioned models used in the samples. couldn't get it to work right at all. actually could never get multiple characters to work correctly. this one was no different

    Saya
    Author
    May 18, 2023· 1 reaction

    The model is CetusMix V2.

    Make sure you installed the LyCORIS additional extension.