CivArchive
    Preview 68087386
    Preview 68087432
    Preview 68087435
    Preview 68087436
    Preview 68087411
    Preview 68087434
    Preview 68087438
    Preview 68087439
    Preview 68087440
    Preview 68087478
    Preview 68087486
    Preview 68087485

    我们需要真正理解女性年龄的模型而不是只能生成萝莉!

    We need a model that truly understands female age, not one that only generates "loli"!

    所以这是一个旨在生成成熟女性并且试图理解多样的女性年龄,而不是只能生成小孩、女孩、萝莉的模型。

    Therefore, this is a model designed to generate mature female and attempts to understand the diverse ages of women, rather than only generating "children, girls, or loli."

    我的推特:@ekll01

    My Twitter: @ekll01

    加入Discord Join Discord

    版本说明请查看About this version

    Please see "About this version" for the version description.

    请参考我公开的工作流

    Please refer to the workflow that I have disclosed.

    目前支持的年龄tag

    女性:lxlx、mesugaki、bishoujo、lady、mature lady

    男性:shxtx、bishounen、gentleman

    请您用以上tag代替1girl、1boy。

    Currently Supported Age Tags

    For females: lxlx, mesugaki, bishoujo, lady, mature lady

    For males: shxtx, bishounen, gentleman

    Please use these tags instead of 1girl or 1boy.

    Recommended Settings(general)

    CFG: 5

    Steps: 40

    Sampling: Euler a

    Scheduler: simple

    Resolution: 832*1216 above

    希望喜欢这个模型的朋友能够量力为模型进行一下Bid投资。这不是请求,但是如果有人能够帮助模型进行Bid,我会十分感谢!

    I hope that friends who like this model can consider making a "Bid" investment in the model according to their ability. This is not a request, but I would be very grateful if someone could help the model with a "Bid."

    Description

    null

    FAQ

    Comments (43)

    yourneighbourRizApr 5, 2025
    CivitAI

    not sure who recommended more oily since most want it removed due to oiliness and generic reflections being burned into most illustrious model by default.
    If anything I'd recommend trying to increase clarity and color saturation a bit, and generate images without hi-res fix to make sure the raw base is clean and stable. If you are mixing in styles like Regura or Raita they heavily desaturate the image and deform features like eyes, hands, etc so it's recommended to add more clean and colorful styles to counterbalance that

    EKLL
    Author
    Apr 5, 2025

    If looking at the comments section and my x comments there is indeed a lot of hope to add glossy information, I should not be misjudging everyone's information.

    Regarding Regura or Raita, I really didn't add those two styles, the dataset was made by me.

    But on the clarity point, I agree, the next step will try to improve this area.

    Thank you for your advice.

    EKLL
    Author
    Apr 5, 2025· 1 reaction
    CivitAI

    If you have any ideas, please leave them in the comments section. I may have entered a closed box, unable to think about the future development of this model outside the existing framework.

    TTangSlgyApr 5, 2025
    CivitAI

    0.6exp问题非常大,脸基本过拟合,的确修脸可以修好,但不开ad,就是过拟合的噪点,的确画面增亮,皮肤增加了油光,反而更像AI出图了,画面亮度可以用slider的lora来调整,油光等也有lora,没必要融入底模,新模型对画师标签基本无效,还不如0.3,作为从0.2过来的人所有版本我都测过,只有0.3最好,我也只保留了该版本,后面版本越来越糟糕,底模只要泛化性好,IL的特点是danbooru标签好用,如果失去这2个,那还不如回PONY呢,是吧。测试了几张图我已经删掉了该版本,说实话我不期待更多版本了,除非是修正0.3的男性年龄或者GROUP情况,否则我以后都懒得下了,0.3已经是该系列最优秀的,不如换个赛道吧,比如noobai,flux等

    EKLL
    Author
    Apr 5, 2025· 1 reaction

    首先0.3版的男性年龄问题是无法被修正的,这个问题源自il0.1的女性和男性外表年龄不一致导致,如果使用比较常见的年龄调整lora的方式,就会发现男性比女性角色更快的老化,这点noobi也不例外。这个版本完全是为了测试新的方式,和0.3进行男性生成就能发现男性老化问题已经初步解决了。目前面部的问题是我制作的数据集出现了问题,我正在重新制作数据集。关于油光问题,中英日语区的需要貌似不太一致,我也比较混乱,因为0.5版本在调低油光比例时又有人要更多油光,我也不清楚后需要如何发展,完全没有可以一起讨论的人。

    TTangSlgyApr 5, 2025

    @EKLL 0.6最大的问题是脸部崩坏以及画师标签(风格)无效,或者说效果很低,看不出效果,因为很多画师的特点不外乎:结构和色彩,结构包括肢体特征和线条,上色就千奇百怪了,一旦增加了过多的颜色比如油光,对比明暗等,就会破坏一些画师标签风格,这是我个人不愿看见的。说实话吧,要求很多的人只是很懒,都希望用一个底模来跑完所有的图并达成想要的效果,然而想满足所有人的愿望是不可能的,对吧。油光明度等都可以用lora后期加,我自己跑图使用7-8个LORA都很正常,但有人就是懒太多这样的就不吐槽了。相反如果融入了太多元素比如油光,想减却没办法,所以底模泛化性才是关键。回到主题,我喜欢你的模型关键就是成熟女性,我认为这也是你的初衷,所以画面感,色彩我认为都是其次并且可调可控的,只要初衷没有改变或者继续增加,其他都是次要的不是吗?(至少我认为在0.6中肢体已经不如以前的更丰满,反而显得过于年轻化,或许也是色感造成的),我认为你的目标应该是成熟女性,成熟男性,年龄可控(向上增加)吧?当然这是我的猜测,作为喜欢该模型的人之一,我希望模型越来越好,或许大家的建议可听,但也可以坚持自己的原则(包括我上面说的一系列废话不喜请无视吧)。总之谢谢你的努力,希望你能更成功。

    EKLL
    Author
    Apr 5, 2025

    @TTangSlgy 画师标签应该就是油光占比太大了,这个倒是比较好解决的,因为我还有一个自己出图使用的版本,那个版本油光占比更小,丰满的话是为了满足更多人的需求将胸部大小之类提示词进行了强化让大家自由控制了,可以看图例组后一张图,这个版本我可能确实把事情想的太复杂了。

    satoshizhouApr 6, 2025

    @EKLL 油光问题让用户自己去折腾就得了,加减几个Lora的事儿,你一个做大模的又不是做保姆。如果说这个模型让你感觉出极限要开新路子了,就另外开个大模的帖子,没必要加分支了,这样你也好做事用户也容易对号入座不混乱。

    EKLL
    Author
    Apr 6, 2025

    @satoshizhou 你说的有道理

    cikodelvin320157Apr 6, 2025

    我之前下过这个老师的版本,但是不知道为啥我相同参数设置,其他大模型能正常出图,这个老师的这个大模型却老是崩溃,虽是如此但也不得不提这位老师的模型熟女的脸是最正的,其他熟女的大模型我跑出来总差点意识,以前的pony的大模型我融合其他Lora也能跑出我喜欢的脸,但是ill的大模型跑出来的图确总是差点意思

    cikodelvin320157Apr 6, 2025

    好家伙,我现在点进你的主页一看你居然是飞燕风Lora的作者,我之前一直用你的Lora跑图,很感谢你提供的作品

    cikodelvin320157Apr 6, 2025

    还有一件事,大佬说的0.3是0.3还是0.3exp版本

    EKLL
    Author
    Apr 6, 2025

    @cikodelvin320157 谢谢您的使用,他说的大概是0.3版本,不带exp,带exp的版本都是实验性质的应该不适合长久用

    TTangSlgyApr 6, 2025

    @cikodelvin320157 说的0.3是不带exp的,包括其他系列我认为效果都不很理想,至于崩溃,如果参考的是PONY的设置,IL使用的是不同的提示词,请注意。@EKLL 我测过许多IL,包括人气高的,但最后的最后,我也只保留你的0.3,因为这是让我最满意的。如前面所言,我认为后期可调可控的元素不适融入底模(特别是容易破坏IL标签的),当然我并非在灌输自己的理念,如果你决定了下一个方向也可以坚持走下去。只是从最初跟随过来,我相信你的初衷是“需要成熟的女性”,所以我认为应该专精这块,画面好的底模太多,然而我认为都太过相似,大家只是追求画面效果,结构脸型几乎一样,我不否认画面效果,然而这些可以靠LORA解决,而我喜欢的你模型的根本也在成熟女性,实际0.3中角色与画师都能很好体现,虽然缺点也一堆,画面是有一些灰暗但基本可用LORA解决。前面我也过于冲动,将来有的新模型我也会下载并测试,我个人而言更希望能专注成熟女性的微调(包括男性结构,毕竟男女搭配干活不累),以产生好的外形(五官及身体)并且不会导致添加LORA而过拟合。能在保留IL标签的情况下微调出个人特点,我认为才是不错的发展方向,当然这些也只是我个人的一厢情愿,不喜请见谅。谢谢你一直的努力,我也会期待新的作品。

    EKLL
    Author
    Apr 6, 2025

    @TTangSlgy 没事没事 我应该会暂时停止这个版本的更新,去做一个noob版本

    cikodelvin320157Apr 7, 2025

    @TTangSlgy 谢谢回复,大佬你的模型我也很喜欢,之前飞燕那个我一直用,但是现在ill模型的不太出画师的风格效果了,请问这个Lora你还更新吗,很期待更新

    cikodelvin320157Apr 7, 2025

    @EKLL 感谢大佬的回复,你的大模型请恕我愚笨弄不太懂只能放弃了,期待您未来的作品,大模型或者lora我都很期待!

    EKLL
    Author
    Apr 7, 2025

    @cikodelvin320157 没事 我之后在更新一个0.6版的简化版就会暂时停更这个模型了,之后会做一个noob的版本,然后公开目前训练的年龄lora和油光lora,在之后再看情况吧

    TTangSlgyApr 7, 2025

    @cikodelvin320157  IL里已经融入很多画师标签(包括角色),比较有名的本子画师我都测试过是包含的,不过某些画师不不太明显也是事实,我有计划对一些“风格不太明显”的做训练(只是计划,我也在观察社区走向和自己时间),但飞燕等比较有名的画师是可以直出的(亲测有效,使用本系列0.3模型),所以目前没有训练的飞燕LORA的安排,或者可以寻找一下其他发布者的,如果生成的结果差强人意,可以尝试更换底模同时请注意使用标签规范及提示词https://huggingface.co/datasets/Laxhar/noob-wiki/tree/main (noobi通用)。@EKLL 我非常期待noobi版本,说实话相对早期的il,noobi更成熟和稳定,而我一直没有转向那边很大程度还是因为我很喜欢你的模型,所以请加油,哪怕不是人人都喜欢熟女,但至少我个人是非常喜欢你的模型。

    EKLL
    Author
    Apr 7, 2025

    @TTangSlgy 谢谢 现在最大的问题是还是女性和男性年龄不一致的问题,男性过老实在不太好解决,我可能会先将重心转移到解决1girl被萝莉和幼女混淆的问题,如果不能解决这个问题,那么就无法同时调整男性和女性的年龄

    506291394340Apr 5, 2025
    CivitAI

    I have a feeling that your model is excellent, but a lot of the loar doesn't fit.

    EKLL
    Author
    Apr 5, 2025

    The experimental nature of this version is relatively large, which may indeed lead to low lora compatibility

    FondantViperApr 5, 2025
    CivitAI

    0.6 exp is very nice so far

    460672106622Apr 7, 2025
    CivitAI

    illuNoob和rouwei基础的模型,它们俩优势和弱势是啥

    EKLL
    Author
    Apr 7, 2025

    优势是强大danbooru标签识别能力和画风再现

    dobomex761604Apr 8, 2025
    CivitAI

    So far I'm seeing no problems with anatomy/hands/fingers, but there's a different "issue": since the characters are oily-looking, they stand out and don't blend into the environments. It looks very strange because the background and all the objects are unaffected by the new style. I suspect that using any "restoration" model might obscure the issue.

    On the upside, this version is very good at generating adult-looking characters by default. Maybe it is worth making non-oily version, it will be the best cougar/milf model IMO - anatomy seems to be the most stable of all other models (and it doesn't need quality tags).

    yourneighbourRizApr 9, 2025

    agreed. if it can replicate the skin texture and style of the image on the right it would be amazing

    EKLL
    Author
    Apr 9, 2025

    @yourneighbourRiz I don't really understand

    yourneighbourRizApr 10, 2025

    @EKLL I was referring to the image on the right side when you open the link (illustriousXL), notice how smooth and painterly the skin is which is unique compared to most other models. A lot of illustrious merges share the same oily effect, bounce light, highlight so there is not much incentive to use one over the other, except for composition and hand quality, backgrounds and lora compatibility.

    EKLL
    Author
    Apr 10, 2025

    @yourneighbourRiz On the right is the illustriousXL original model, this feature will disappear after adjusting the age, I'm not sure how to fix it...

    SashaHostynApr 8, 2025
    CivitAI

    Hi EKLL,

    I’m getting a recurring NansException: A tensor with NaNs was produced in Unet error when using

    your 0.6 exp model, even after trying:

    1. Adjusting resolutions (tested 1920x1080 and lower)

    2. Tweaking CFG (3-6) and steps (20-35)

    3. Disabling "half precision" optimizations for both UNet and VAE

    My RTX 4070 Ti Super handled fine at 1920x1080 at beginning, but it got crashed after generating several images. VRAM usage sometimes peaks to ~10/16GB and outputs a black image; other times crashes with VRAM overflow.

    Any ideas? Could this be model-specific? Appreciate any troubleshooting tips!

    EKLL
    Author
    Apr 8, 2025

    euler_ancestral_cfg_pp

    cfg: 1

    This is the best setting (personal experience)

    I've had crashes. I'm trying to figure out why

    EKLL
    Author
    Apr 8, 2025

    euler_ancestral_cfg_pp

    cfg: 1

    This is the best setting (personal experience)

    I've had crashes. I'm trying to figure out why

    skindentation321Apr 8, 2025
    CivitAI

    i cant get v0.5 to work at all. it just bricks and gives me a black screen every time. all my other models work totally fine.

    EKLL
    Author
    Apr 8, 2025

    comfyui or webui?

    YagokoroEirin8556Apr 9, 2025
    CivitAI

    always wrong color in comfyui

    EKLL
    Author
    Apr 9, 2025

    Use my workflow?

    EKLL
    Author
    Apr 9, 2025· 1 reaction

    @nam12122003831 This model is inclusive of vae

    luguxunmengApr 10, 2025

    @EKLL 我猜是CFG的问题,您提供的CFG1太小,传统的CFG7过大,我目前使用是CFG4+0.9降噪

    EKLL
    Author
    Apr 10, 2025

    @luguxunmeng 是euler_ancestral_cfg_pp的情况下吗

    luguxunmengApr 10, 2025

    @EKLL 是euler_ancestral,euler_ancestral_cfg_pp+cfg1似乎倾向于生成单调背景(相关prompt效果也一般)

    EKLL
    Author
    Apr 10, 2025

    @luguxunmeng 对这个模型来说大家按个人习惯使用采样器就行,Vpred版本就必须要遵从官方要求了

    Dgx09Apr 9, 2025
    CivitAI

    seems to be missing the ladies, make up, tatoo, ...

    Checkpoint
    Illustrious
    by EKLL

    Details

    Downloads
    1,315
    Platform
    CivitAI
    Platform Status
    Available
    Created
    4/5/2025
    Updated
    5/14/2026
    Deleted
    -

    Available On (1 platform)

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