CivArchive
    Preview 136003285
    Preview 136003282
    Preview 136006689
    Preview 136006657
    Preview 136006678
    Preview 136003293
    Preview 136006691
    Preview 136006658
    Preview 136003287
    Preview 136006669
    Preview 136006663
    Preview 136006666
    Preview 136006668
    Preview 136006676
    Preview 136006674
    Preview 136006675
    Preview 136006687
    Preview 136006682
    Preview 136006680

    喵喵后宫才不是臭脚模型呢

    属于完整的 MiaoMiaoHarem Dogma 完整生态系统终于来了。👈详情请点击

    👑 MiaoMiao Harem Anima | 二次元美学新纪元!基于 Anima 架构的华丽降临

    “不是简单换底模,而是让喵喵后宫的审美灵魂在 Anima 架构上完成真正二次元觉醒!” MiaoMiao Harem Anima 正式发布!本次更新基于 Base 1.0 版本深度迭代:

    弱化过度风格化,保留并强化核心美感默认风格更自然,构图质量大幅提升更好延续了喵喵后宫系列一贯的审美与构图特点

    我们以 Anima BASE1.0 架构为核心,打造更纯正的二次元表现力。相比泛用写实底模,Anima 在动漫语义理解、角色风格稳定性和提示词兼容性上拥有先天优势。核心亮点:

    提示词理解极度友好:纯标签、自然语言、混合写法均可轻松驾驭出图高效:简单 prompt 即可获得高完成度,易控且上限高定位清晰:专注动漫插画、非写实艺术表现,强项在于角色立绘、幻想场景、轻小说封面、游戏卡面等二次元审美

    这不是普通的模型分支,而是喵喵后宫在二次元赛道上的全新主场版本——更纯、更稳、更懂你。喜欢浓郁风格化的朋友也不用担心,后续将推出基于 Anima 1.0 的强风格化版本,敬请期待! 立即体验,开启你的动漫幻想盛宴!欢迎分享返图~

    目前模型训练完全由我个人承担,资源有限I’m personally funding all model training with limited resources.

    为了让大家更早体验更多内容,恳请支持!Your support helps more people experience new features sooner!

    您的赞助将用于 GPU 费用与模型优化。All contributions go toward GPU costs and model improvements.

    感谢您的鼓励与信任!Thank you for your kindness and trust!

    https://ko-fi.com/miaokamk Please sponsor me. Thank you for your support!

    https://afdian.com/a/MIAOKA 请您赞助我 感谢支持 !

    ⚠️IMPORTANT: Usage Guide / 重要食用指南

    请查阅各版本下方的「About this version」区域,以获取详细的使用说明与推荐参数(如图所示)

    ⚡ 二次采样优化 / Secondary Sampling Enhancement:

    高清放大不糊不炸,细节质感全面觉醒!

    针对 Anima 模型 在高分辨率二次放大过程中容易出现的 模糊、发花、线条脏乱、细节崩坏 等问题,我在 WebUI Forge Neo 中反复实测,终于调出了一套更加均衡稳定的二次采样参数方案。 这套设置可以在尽量保留原图角色结构、神态、构图与插画气质的同时,显著提升整体清晰度与细节质感——放大后不容易糊,不容易花,线条更稳,画面更干净,成图完成度更高。无论是:•角色立绘•半身插画•卡面封面•壁纸级高分辨率输出•细节要求更高的成稿场景都能获得更舒服的放大效果。

    直接按照下图参数操作即可,基本就是当前 Anima 版本较为平衡、稳定、实用的二次采样方案。

    有关如何使用 V-pred 模型的说明,请参阅 ←点击

    您可以通过加入我们的 Telegram 和 QQ 群组获取其他版本抢先体验资格,与我们一起见证这次重大的更新。感谢大家一直以来的支持和反馈!我将继续努力,为大家带来更多精彩的更新和功能。

    ☜ 请您赞助我 感谢支持 !Please sponsor me. Thank you for your support! 您的赞助将用于 GPU 费用与模型优化 All contributions go toward GPU costs and model improvements


    ⬅️企鹅1群 (目前已满)⬅️Telegram

    ⬅️企鹅2群(已满)

    ⬅️企鹅3群(开放中)


    这一次,我们选择以 Anima 2B 架构为核心,重新塑造喵喵后宫的二次元表现力。 不同于偏向泛用或写实混合取向的底模,Anima 天生就是为 anime / illustration / artistic image 而生——它拥有更纯正的二次元语义理解、更稳定的角色风格表达,以及对标签体系与自然语言混合提示更强的兼容能力。

    基于这一点,MiaoMiao Harem Anima 不再只是“能画二次元”,而是向着更纯净的动漫审美、更鲜明的角色塑造、更统一的插画质感全面进化。 从角色立绘、单人美少女、双人互动,到轻剧情感场景、幻想服装设计、插画封面感构图——都能以更自然、更轻松、更高完成度的方式直出。

    这不是一次普通换底,而是一次真正面向二次元创作者的专属升级。

    提示词理解更友好 / Better Prompt Understanding:

    标签、自然语言、混合写法都能打

    Anima 原版就具备一个非常强的优势: 同时支持 Danbooru 风格标签、自然语言描述,以及二者混合输入。

    因此在这个版本中,你可以更自由地使用不同提示习惯:

    纯标签流:适合老手、控图党、精确指定元素

    自然语言流:适合描述氛围、构图、服装、剧情感

    标签 + 英文描述混合流:兼顾稳定性与创造力

    这意味着它既适合喜欢堆 tag 的用户,也适合只想用几句自然语言快速出图的用户。 对于角色名、作品名、服装设定、场景描述、画师风格倾向等内容,也能更自然地接收与还原。

    一句话总结: 更懂你想要什么,也更容易把“你脑子里的二次元画面”真的画出来。

    ⚡. 出图效率与可控性兼得 / Fast, Flexible, Effective

    虽然这是一个偏基础能力很强的动漫架构模型,但它并不意味着难用。 恰恰相反,MiaoMiao Harem Anima V1.0 保留了喵喵后宫系列一贯的低门槛、高回报特性:

    •简单 prompt 也能获得不错完成度

    •加质量标签后上限进一步提升

    •通过角色/服饰/姿势补充可快速细化画面

    •适配多种采样器,出图风格变化明显

    •可用于轻量控制,也可用于深度调校

    无论你是想“一抽出图”,还是想认真打磨构图与细节, 这个版本都能在效率审美稳定性之间取得非常优秀的平衡。

    非写实路线更加明确 / Illustration-Oriented by Design:不追求现实,只服务于幻想与艺术

    需要特别说明的是: MiaoMiao Harem Anima V1.0 是一个明确偏向动漫插画与非写实艺术表达的版本。

    它的强项不是现实摄影感,也不是高仿真写实人像, 而是:

    •动漫立绘•二次元角色插画•游戏卡面感视觉•轻小说封面感构图•梦幻幻想风•艺术化非写实场景

    如果你追求的是“像动画、像插画、像角色原画、像设定集封面”的视觉方向, 那么这个版本会比偏写实底模更加纯粹、更加直接、也更对味。

    MiaoMiao Harem Anima V1.0 不只是“二次元角色模型”, 更是一台可以在动漫、插画、艺术表现之间自由摆动的创作引擎。

    MiaoMiao Harem Anima V1.0,不是简单分支,而是喵喵后宫在二次元赛道上的全新主场!

    它继承了喵喵后宫系列的高审美表达,同时借助 Anima 架构获得了更纯正的动漫理解、更舒服的插画完成度,以及更友好的 prompt 兼容能力。 对于热爱角色创作、动漫插画、卡面海报、幻想美少女题材的创作者来说,这将会是一次非常值得体验的新进化。

    如果你想要的是更纯、更稳、更有“二次元味道”的喵喵后宫版本——那么这就是答案。

    立即体验,开启属于你的动漫幻想盛宴!欢迎分享返图,一起见证 Anima 架构下的全新蜕变!

    这个模型如你们所见,由于我的个人兴趣,我更热衷于画脚。但是,如果你认为这是一个脚部特化模型,那就错了。这个模型的特点不在于画脚,而在于每个版本都在优化质感、语义理解和皮肤质感。我想创造一个能画出类似写实肌肤的动漫风格为主体,并尽量不影响其他风格的模型。这个模型可能很小众,没有多少人使用,但我会继续更新迭代以达到我的目标。如果你们喜欢,请帮我赞助我 BUZZ,帮忙点点赞、多评论、多返图就好了。

    As you can see, this model reflects my personal interest in drawing feet. However, if you think this is a foot-specialized model, you're mistaken. The model's strength lies not in drawing feet, but in optimizing texture, semantic understanding, and skin texture with each version. My goal is to create a model that can produce anime-style art with realistic skin textures while minimizing the impact on other styles. This model might be niche and not widely used, but I will continue to update and iterate to achieve my goal. If you like it, please support me on BUZZ by liking, commenting, and sharing more images.

    Model Usage Statement

    • 1. You are not allowed to host this model and its derivative versions (such as fusion model versions) on websites/applications that plan to earn income or donate. If you need to use this model and its derivative versions for commercial purposes (such as generative services, selling images, using images for public publication on short video platforms, etc.), please contact me through the website

    • 2. You are not allowed to directly sell images generated by this model, unless you have made sufficient manual modifications to the image so that it can be fully recognized as your personal work in a legal sense. If you violate this provision, all legal consequences resulting from it shall be borne by you personally, and I am not responsible.

    • 3. You cannot use this model to intentionally create or share illegal or harmful content for dissemination and output. Please abide by public order and good ethics, and

    Description

    Update Highlights / 更新亮点

    MiaoMiao Harem Anima V1.4 现已正式更新!

    本次 V1.4 是一次聚焦“风格辨识度”的关键迭代。在 V1.2/V1.3 持续探索并确立自然柔和与画面张力平衡点的基础上,我们进一步调整了模型的风格化权重,核心目标明确:让模型真正拥有属于自己的美学签名。毕竟,若一个衍生模型缺乏鲜明的特色表达,便失去了独立存在的意义——与其平庸地贴近原版,不如自信地放大系列独有的审美主张。

    V1.4 主要更新内容: ● 风格权重精准调优:在保持底层画面流畅自然的前提下,适度拔高风格表现力,强化线条张力、色彩氛围与角色神态的独特性,使出图更具“辨识度”与喵喵后宫系列专属的艺术韵味 ● 基础结构日常迭代:底层网络架构持续微调优化,进一步提升模型稳定性、泛化能力与细节还原精度,崩坏率维持低位,复杂场景适配更从容 ● 提示词体系轻量延续:正负向提示词保持 V1.3 的简洁高效设计,无需冗长堆砌,简单自然的 prompt 即可稳定触发高品质出图 ● 工作流无缝衔接:推荐采样器、步数、CFG 及分辨率等核心参数与 V1.3 完全一致,老用户可直接沿用原有配置,零成本平滑升级

    在训练过程中,我们围绕“特色鲜明”与“泛化稳定”进行了多轮权重校准与结构验证。每一次调整都力求让模型既敢突出个性,又不失底层逻辑的扎实可靠,只为更精准地释放 Anima 架构的潜力。

    定位延续:依然专注于纯正二次元插画风格,强于高品质角色立绘、幻想服装、轻小说封面感构图与梦幻美少女场景。V1.4 将进一步放大系列的高审美基因,为创作者提供更鲜明、更具感染力的视觉表达工具。

    感谢每一位持续陪伴与反馈的用户。V1.4 是一次“回归特色、强化个性”的迭代,希望它能为你带来更契合预期、更具创作灵感的出图体验!

    ⚠️ IMPORTANT: Usage Guide / 重要食用指南

    Model Output Parameters
    Scheduler: Euler Normal / Euler a Normal
    Steps: 30
    CFG Scale: 4.0 - 5.5
    Recommended Resolution: 920x1536

    Positive Prompt(默认模板)

    best quality, score_7, score_9, sensitive, very aesthetic, ultra detailed,fair skin, explicit, high contrast, {此处添加提示词 Add Prompt here}
    

    Negative Prompt

    worst quality,  low quality,  score_1,  score_2,  score_3,  artist name
    

    💡 提示:负向提示词可根据具体需求灵活补充,但建议保持精简。当前版本已大幅降低对冗长提示词的依赖,简单自然的 prompt 即可达到极佳效果。

    Thank you for your continuous support and feedback! We hope V1.4 brings you a more distinctive and inspiring creative experience.
    感谢每一位持续陪伴与反馈的用户。V1.4 是一次“回归特色、强化个性”的迭代,愿它能为你带来更契合预期、更具创作灵感的出图体验!

    FAQ

    Comments (33)

    goeypants400Jul 7, 2026· 10 reactions
    CivitAI

    absolute GOAT who still doesnt use early access on this model

    rastandaaJul 8, 2026
    CivitAI

    is there anyhow we can know how updated the character used by the training for each version?

    Sakura_leiyingJul 8, 2026
    CivitAI

    1.4的色彩有些发灰,完全相同提示词,使用该模型,白色背景变成灰色,颜色也不够鲜艳,作者是刻意如此,还是未达预期

    MIAOKA
    Author
    Jul 8, 2026

    刻意

    MIAOKA
    Author
    Jul 8, 2026

    原本风格就该如此只是回归了1.2版本之前的版本风格就是这样的

    MIAOKA
    Author
    Jul 8, 2026· 2 reactions

    你可以看这个版本简介我有提到 我这版要突出风格 如果一个模型没有独特的风格就不如用原版模型了

    MIAOKA
    Author
    Jul 8, 2026· 6 reactions
    CivitAI

    询问一下大家的意见: 各位是否需要 Anima 模型的 int8 或 FP8 量化版本?如果需求的人比较多,我会考虑发布一个量化版。 不过,原版 Anima 模型本身对配置要求并不高,我也在犹豫是否有发布的必要。同时想提醒大家:如果没有相应的加速环境(如 TensorRT、ONNX 等),使用 int8 或 FP8 量化版本时,出图速度可能会大幅下降(成倍变慢),图像质量也会因量化而有所降低。 欢迎大家留言反馈,谢谢!

    SimonFranck1992Jul 8, 2026· 3 reactions

    在这里不太好说,如果有人说不需要,那就有人肯定喷:你不要别人要。但如果有人说要,就会又有人喷:Anima本身就不需要多好的配置。注定会在这里吵起来,目睹过太多这种事了,个人感觉还是私信比较好,两拨人互不干涉。

    yumimikoJul 9, 2026· 1 reaction

    个人觉得还是把资源集中在更新模型上

    MIAOKA
    Author
    Jul 8, 2026
    CivitAI

    Asking for your opinions:Does everyone need an int8 or FP8 quantized version of the Anima model? If there is sufficient demand, I’ll consider releasing one.However, since the original Anima model already has relatively low hardware requirements, I’m also wondering whether it’s necessary. Please note that without the proper acceleration environment (such as TensorRT, ONNX, etc.), using int8 or FP8 quantized versions may result in significantly slower generation speeds (potentially several times slower) and a noticeable drop in image quality due to quantization.Feel free to leave your feedback below. Thank you!

    spstavrakis988Jul 8, 2026· 3 reactions

    i mean do whatever you are comfortable with doing

    BesssJul 8, 2026· 4 reactions

    GTX 1650 user here, the idea is pretty appealing, and i do think that it's possible to make it work, but it'll take too much setup the model, it wouldn't be the best option unless the user initally already required to use a "proper accelerated runtime" to even fit the normal model without spilling it over their VRAM. not to mention quantization might cause issues with loras and stuff. So.. I think a quantized model would only be useful for a very small minority.

    [Note : This opinion is based on my vague understandings, it might contain mistakes, if there is, feel free to correct me]

    AlternativeAlterJul 9, 2026· 1 reaction

    As a very low VRAM user (Just 4GB on a laptop GPU), I can run the base fp16 model just fine in comfortable speeds and personally, even if the higher quant version was realeased, I wouldn't bother with setting the whole environment just for a bit of extra speed so I wouldn't use it. The requirements for running it are already so low, that unless someone is runnig it on a potato, higher quants aren't needed. So In my personal experience, I don't believe that FP8 or int8 are needed, especially if it's extra work that could be better used somewhere else. But it's just my five cents.

    BTW, really amazing model, I'm absolutely having a blast with it :)

    spstavrakis988Jul 9, 2026· 1 reaction

    @Besss yeah these small models really don't need quantization, since artifacting is possible with quantization and, most consumer cards dont mathematically benefit from fp8 models

    thalesasr714Jul 9, 2026· 1 reaction

    Even if the model runs fine in FP16, I feel like a int8 convrot would be awesome. More than doubling speeds with barely any image quality loss would be nice. I'd skip fp8 and int8 imo, since apparently int8convrot seems to be just plain superior in both performance and compatibility of older video cards (RTX 20 and older).

    NeonCheckpointJul 8, 2026· 3 reactions
    CivitAI

    This is singlehandedly one of the best checkpoints I've ever used!

    这绝对是我用过最好的检查点之一。!

    ZerovbJul 9, 2026
    CivitAI

    I wanted to try it but it is not enabled :c

    zjc1772665101270Jul 9, 2026
    CivitAI

    Do the CLIP strength and LoRA strength need to be kept the same when using a LoRA? Thanks.

    3315213393920Jul 9, 2026
    CivitAI

    发现控制降噪的步数,调整在10的时候就会出现非常朦胧的画风。但是不影响质量

    JianEthanJul 9, 2026
    CivitAI

    我发现这个文本编码器好像不一样,我用的是qwen_3_06b_base,效果一样吗?

    MIAOKA
    Author
    Jul 9, 2026

    一样

    1296515195758Jul 10, 2026
    CivitAI

    Insane versatility. This LoRA handles everything from close-ups to wide shots effortlessly

    zyk262Jul 11, 2026
    CivitAI

    期待加速模型

    pnsyfeqfJul 11, 2026
    CivitAI

    absolutely god

    samson4617Jul 13, 2026
    CivitAI

    When using image with close-up shot for img2img without action prompts, the chance for obtaining lewd outputs is very high. 有嘗試把'safe'加到提示詞中但是並沒有太多作用🤪

    NullNameAlterJul 14, 2026

    可以把【Exposed spot,】加入负面提示词,就可以完美避免掉

    MIAOKA
    Author
    Jul 13, 2026· 18 reactions
    CivitAI

    The experimental "8-Step Distill" version of MiaoMiao Harem Anima has entered internal testing and will soon be available to everyone.

    MiaoMiao Harem Anima 的实验性“8步蒸馏”版本已进入内部测试阶段,很快就能和大家见面。

    zjc1772665101270Jul 15, 2026

    great

    dekmon231576Jul 13, 2026
    CivitAI

    我是新手,看示例图里有的加入了角色名字,有的没有加。后者是怎么实现角色还原的呢?就光靠服装和外貌提示词吗?

    tim7799822208Jul 21, 2026

    去AnimaDEX網站看

    ecchosamaJul 14, 2026
    CivitAI

    RuntimeError: ERROR: clip input is invalid: None

    hn87Jul 15, 2026

    @ecchosama, to run Anima-based models you need to download external text encoder and VAE https://huggingface.co/circlestone-labs/Anima/tree/main/split_files


    bambiettaJul 14, 2026· 13 reactions
    CivitAI

    虽然个人风格比较强,但这应该是美学水平最高的一个anima微调模型,不需要其他复杂的lora和提示词就能让生成的图像质量水平很高👍

    Checkpoint
    Anima

    Details

    Downloads
    11,017
    Platform
    CivitAI
    Platform Status
    Available
    Created
    7/7/2026
    Updated
    8/4/2026
    Deleted
    -

    Files

    miaomiaoHarem_anima14_txt.safetensors

    Mirrors

    HuggingFace (96 mirrors)
    CivitAI (134 mirrors)
    ModelScope (1 mirrors)

    qwen_image_vae.safetensors

    Mirrors

    HuggingFace (211 mirrors)
    CivitAI (146 mirrors)
    TensorFiles (1 mirrors)