CivArchive
    MiaoMiao Harem - Anima_1.1
    NSFW
    Preview 132785808

    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与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

    About this version

    MiaoMiao Harem Anima 1.0 已上传

    这个版本是去风格化的 MiaoMiao Harem Anima bsae的基础版本的加练版本

    出图参数延续

    ⚠️ IMPORTANT: Usage Guide / 重要食用指南

    Model Output Parameters

    Scheduler Euler Normal Euler a Normal

    Steps 30

    CFG Scale 4.0-5.0

    Model Recommended Resolution 920x1536

    Positive Prompt （default ）

    masterpiece, best quality, score_7,  score_9,safe,very aesthetic, ultra detailed,pale skin,fair skin, 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,simple background,lowres,(bad),text,error,fewer,extra,missing,worst quality,jpeg artifacts,low quality,watermark,unfinished,displeasing,oldest,early,signature,artistic error,username,scan,[abstract],english text,shiny hair,extra toes,

    ⚡ 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。

    直接按照下面蓝字链接的参数操作即可，基本就是当前 MiaoMiao Anima 版本较为平衡、稳定、实用的二次采样方案。二次采样参数设置图

    FAQ

    Comments (43)

    ppap11404Jun 4, 2026· 3 reactions
    CivitAI

    我非常喜欢这个风格，不过目前有两个常见问题一直困扰我，其一是手部经常出现六指，反复抽种子过程中依然是六指概率非常高，其二是在出角色站立全身图的时候，视角会倾向于从上往下的广角，很难用提示词修正

    MIAOKA
    Author
    Jun 4, 2026· 1 reaction

    anima系列我正在努力修复手脚肢体问题 关于视角你可以拉高视角提示词权重就能解决或者使用自然语言来描述视角

    YabbaJun 4, 2026· 2 reactions
    CivitAI

    i dont know if this is just a me thing, but have any of you all had any trouble trying to get different views for an image, like above, below, from the side? felt like it genuinely would not listen as if the prompt didnt exist which is a first for me ever.

    MIAOKA
    Author
    Jun 4, 2026· 1 reaction

    Regarding perspective, you can solve it by increasing the weight of the perspective prompt, or by using natural language to describe the perspective.

    lkorint386Jun 5, 2026· 2 reactions
    CivitAI

    模型生成的人物总是占满画面，像是怼脸拍大头照，几乎不会有很多留白，是anima的问题吗佬

    GPUPoorChadJun 5, 2026· 3 reactions
    CivitAI

    Really great default style, the last version would ignore artist tags but this one seemingly has the best of both worlds somehow. Might be my favorite Anima checkpoint so far, WAI is close 2nd. However, one thing I have to ask is why do the example prompts use underscores? To my understanding Anima is intended to be used with spaces and escape characters for ( etc. Is it trained with underscores? If so I think this is bad design choice personally, but I won't pretend to be an expert just feels like it would be best to train most like how it was originally trained by Circlestone

    MIAOKA
    Author
    Jun 5, 2026· 2 reactions

    Here's the thing. I'm using the Neo branch of webui with the Anima model to generate images. Today, someone mentioned the issue with underscores. Since I'm using webui neo, I installed a plugin to organize prompt word structures, which automatically converts some spaces in prompt words to underscores. However, these prompts are not effective because Anima only recognizes underscores similar to those used in Pony aesthetic scoring. Other underscores should be converted to spaces. I will make the necessary changes later and recommend quality tags

    Nns8izJnE67nJun 5, 2026· 1 reaction

    Scores specifically in Anima use underscores.

    ChenkinJun 5, 2026· 9 reactions
    CivitAI

    这正是我理想中的anima模型

    shenhaixianyuJun 5, 2026· 3 reactions
    CivitAI

    十分喜歡這個模型的，但使用 hires fix 放大 1.5，按建議 denoise 0.8 會使人物動作改變，畫面重繪相差太遠。denoise 0.5 發花太嚴重，有時候會有黑點出現。怎麼辦呢？

    MIAOKA
    Author
    Jun 5, 2026· 1 reaction

    Anima架构限制 只能是这样了 如果你用的 webui neo分支 按照我的二采设置是没问题的 如果你是用comfy ui 不能按照我neo的设置来使用 comfy ui有相关更稳定一点的 放大节点 和anima成品带放大节点方案的工作流 效果是优于 我webui neo分支推荐的二次采样放大的方法的

    shenhaixianyuJun 5, 2026· 1 reaction

    有找到 anima segs upscale 的工作流，的確能用。只不過把圖分區重繪真的很慢，而且為了無縫合併，denoise 只能開 0.35，細節增加有限，亦不能使用 detail demon sampler，有點太可惜了。莫非 anima 在圖大於某 size 就會開始發花？而分區卻不會？搞不太懂🫠

    MIAOKA
    Author
    Jun 5, 2026· 1 reaction

    @shenhaixianyu 是的他不能像SDXL架构 一样 直接二采放大 二采放大倍率超过1.5倍就会开始画图崩坏 个人认为 anima二次采样对于噪声比较敏感 我推荐的neo的二次采样 也是卡在临界值放大 效果也不是很好 但是能用

    MIAOKA
    Author
    Jun 5, 2026· 1 reaction

    @shenhaixianyu 主要现在anima的 ControlNet 还没有 tlie的模型 如果后期有基于anima的tlie ControlNet 模型后 基于tlie放大应该是不错的选择

    shenhaixianyuJun 5, 2026· 1 reaction

    祈望 anima 的團隊可以發現問題，從底層源頭解決就最好了。 🙏

    MIAOKA
    Author
    Jun 5, 2026· 1 reaction

    @shenhaixianyu 貌似不太可能 因为从预览版就存在了 现在已经是正式版了应该不会改变了 现在只能靠之后有没有其他的通过后期的方法来解决了

    cpufanJun 6, 2026· 1 reaction

    因为不同的采样器受denoise参数的影响程度不同。

    KOOOCJun 5, 2026· 3 reactions
    CivitAI

    适用于comfyui吗

    MIAOKA
    Author
    Jun 5, 2026· 1 reaction

    适用于Comfy ui 如要工作流可以加群 我群内有工作流

    Gibson1337Jun 5, 2026· 1 reaction
    CivitAI

    i get just noise and im using all the correct vae, encoder, and reccomended settings.

    MIAOKA
    Author
    Jun 5, 2026· 1 reaction

    What UI are you using? Is it WebUI, WebUIForge, or ComfyUI? If it's WebUI or WebUIForge, it cannot be run directly. WebUI requires the Neo branch of WebUIForge to run. ComfyUI can be run using the correct workflow.

    heywhatJun 6, 2026· 4 reactions
    CivitAI

    为啥质量提示词还带一个score_7啊？融pony了？

    MIAOKA
    Author
    Jun 6, 2026· 3 reactions

    因为anima base原版训练的时候就沿用了pony模型作者的美学评分 所以就可以加pony标签 不是你想象的融的pony 这两个模型一个dit架构一个unet架构 架构都不同 权重key都对不上 融完全是不可能的事

    RivaleJun 6, 2026· 3 reactions
    CivitAI

    please help me, I use A1111, but gray squares constantly appear during generation. there was no such thing with IL models

    MIAOKA
    Author
    Jun 6, 2026· 1 reaction

    Because the webUI, specifically A1111, hasn't been updated in a long time and is no longer compatible with Anima, you either need to use comfyui or the webUI forge neo branch to use it.

    assasingame549439Jun 8, 2026· 1 reaction

    @MIAOKA I do use confyUI and it occured to me, that clip skip 2 leads to that. How do you use clip skip with that text encoder?

    Cimra11Jun 8, 2026

    @assasingame549439 thank you for telling me what was the issue

    assasingame549439Jun 10, 2026

    @Cimra11 Glad if i manage to help you somehow. Btw i discovered that ddim sceduler with dpms_2s_ascentral sampler works way better than recommended one

    frigidblast381Jun 6, 2026· 3 reactions
    CivitAI

    the best one so far, thanks boss!

    tsukfhgd244Jun 6, 2026· 2 reactions
    CivitAI

    请问，不同大模型备好的提示词模块，

    用在miao miao harem上，

    效果不一样吗，

    在别的大模型上，适配出来品质很高的图，

    但是一样用在miao miao harem上，

    感觉品质就不太行，

    但是我好喜欢miao miao harem的画风。

    SissyleonaJun 7, 2026· 2 reactions
    CivitAI

    Urgh i'm sorry for the ranting comment...

    But AI is so hard. (Mind you i'm a huge noob and I'm heavily struggeling finding info or understanding the info i'm getting so far. For that the way I describe things might be a bit off...)
    So i'm used to illustrious AI checkpoints and loras, which i used with the AI webui... I believe it's AUTOMATIC1111?

    Now I read everywhere that ANIMA is better and bigger and I see many of the loras i use update to it... I read i need to install "Stable Diffusion WebUI Forge - Neo" to use ANIMA checkpoints... so i did... but now i must say... the image generation doesn't follow my prompts that well... it only works when i add a VAE and there are only 2 anima VAE's on civitAI??? And erghhh i'm so bloody confused.

    krade07Jun 7, 2026· 3 reactions

    If you want an easy solution I really like swarmui. You can install it through stability matrix which is a package interface that lets you install comfyui, a111, forge ect. When you try to generate an image it will just download the vae and text encoder automatically. You might need to use some quality tags at the beginning of your prompt like masterpiece, best quality, score_9, score_8, score_7, absurdres, ultra detailed. You might need to try a different scheduler and sampler. Start with euler/ normal. CFG 6, Steps 30-40. Some models are different though. Use the download pages for best settings that the creator came up with.

    frigidblast381Jun 7, 2026· 2 reactions

    automatic1111 is so outdated, if you still want to use webui like, use the Webui Forge-Neo, do it now.

    behemoth1887Jun 7, 2026· 1 reaction

    I think you didn't get the VAE issue.
    You only need two files, qwen_image_vae for VAE and qwen_3_06b_base for text encoder, both can be found on hugging face, then put them in the corresponding folder.
    When first time to load the files, you will find VAE and text encoder files share the same UI slot, don't worry and load qwen_image_vae first then qwen_3_06b_base.
    Finally, different from illustrious, all the quality prompts should be put in the first line, then are the other prompts; or else your output images would be broken.

    SissyleonaJun 7, 2026· 1 reaction

    @krade07 Thank you very much for commenting! This is very helpfull! I'll look into it. Crazy how searching so much on google and only now I learn about swarmui.

    SissyleonaJun 7, 2026

    @behemoth1887 OH!! So the order of all the tags, quality and image description tags matter???
    I didn't know at all.

    So first the quality prompts like : "masterpiece, best quality, score_7, ETC..." and then follow it up with my personal stuff and loras?

    behemoth1887Jun 7, 2026· 1 reaction

    @Sissyleona Yeah, you just can't stack all the prompts randomly anymore.

    krade07Jun 7, 2026· 1 reaction

    @Sissyleona Yeah, swarmui is great. I'm surprised more people don't use it. Also its an overlay for comfyui so if you need to you can just load someones comfyui workflow into and you're all good. Also yeah the prompt would be your quality tags that I typed (those are for anima btw, other models can be different) and then something like dynamic lighting, 3d, rendered, 1girl, blonde hair, blue eyes, red dress, thigh highs, red high heels, black choker, earrings, luxurious ballroom, windows, moonlight, chandeliers. You can get more specific to. Anime is pretty good at doing what you want. Hope it helps.

    SissyleonaJun 8, 2026· 1 reaction

    @krade07 It really helps! Thank you so much for the info and taking your time to help me! I wish you the best of days! <3

    FrankainsteinJun 10, 2026· 3 reactions

    (dropping in case if you still need help ! ) I use forge NEO , you can use it you just need to have the vae for it work and also switch to anima mode there is like a little menu with writted XL,flux etc when you click on it, it open a list of mode you choose anima, then you also gotta have qwen_image_vae.safetensors qwen_3_06b_base.safetensors those two one is your vae , one is your text encoder why because this models can also understand text not just tags like illustrious so it require a bit more stuff if you get an error you can always unlod the vae and try again, but it should not be too hard just chill and try stuff get the right files a Anima checkpoint those vae, switch to Anima mode, for starting use 15 step, shift 3, cfg (as you want but usually i got between 1 and 4) then go at it ! look for the VAE on hugging face or google if you don't find it here but i think they are on civit ai both the vae and text encoder, personally I use stability matrix i would recommend installing it if you are on window and use NVdia it can install Forge neo and comfy ui or any interface you want do a lot of the job for you !

    behemoth1887Jun 7, 2026· 1 reaction
    CivitAI

    If I feel the image is too bright, is it possible to fix it using prompts? What kind of prompts should I use?

    cozenJun 7, 2026· 5 reactions
    CivitAI

    Love this model. I've tried over 100 models from this site. This is the best anime model there is to date.

    Aura_farmmerJun 7, 2026· 1 reaction

    How

    Checkpoint
    Anima
    by MIAOKA
    Download (Beta) View on CivitAI
    anime
    sexy
    girls
    base model

    Details

    Downloads
    5,317
    Platform
    CivitAI
    Platform Status
    Available
    Created
    6/4/2026
    Updated
    7/26/2026
    Deleted
    -

    Files

    miaomiaoHarem_anima11.safetensors

    Size:
    3.89 GB
    SHA256:
    cb62e479a986ce45de48b2a507eb0a1bcc045d2393c046122000878e5738dacd

    Mirrors

    HuggingFace (2 mirrors)
    CivitAI (2 mirrors)

    miaomiaoHarem_anima11_txt.safetensors

    Size:
    1.11 GB
    SHA256:
    cd2a512003e2f9f3cd3c32a9c3573f820bb28c940f73c57b1ddaa983d9223eba

    Mirrors

    HuggingFace (89 mirrors)
    CivitAI (116 mirrors)
    ModelScope (1 mirrors)
    ModelScope CN (2 mirrors)

    miaomiaoHarem_anima11.safetensors

    Size:
    3.89 GB
    SHA256:
    cb62e479a986ce45de48b2a507eb0a1bcc045d2393c046122000878e5738dacd

    Mirrors

    HuggingFace (2 mirrors)
    CivitAI (2 mirrors)

    miaomiaoHarem_anima11.safetensors

    Size:
    3.89 GB
    SHA256:
    604b99e0d9ad18cb648043d62b53ed64d0c8f9238accf043602298687741213a

    Mirrors

    CivitAI (1 mirrors)

    qwen_image_vae.safetensors

    Size:
    242.05 MB
    SHA256:
    a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f

    Mirrors

    HuggingFace (190 mirrors)
    CivitAI (125 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (3 mirrors)
    ModelScope CN (5 mirrors)