CivArchive
    Preview 136673313
    Preview 136673312
    Preview 136673319
    Preview 136673320
    Preview 136673326
    Preview 136673328
    Preview 136673331
    Preview 136673338
    Preview 136673336
    Preview 136673340
    Preview 136673339
    Preview 136673341
    Preview 136673343
    Preview 136673342
    Preview 136673346
    Preview 136673344
    Preview 136673348
    Preview 136673347

    I just made a new Telegram Group and Discord group drop in for chats, questions, feedbacks or just to share your work. 欢迎加入纸飞机Discord群组。

    RH体验地址(邀请码Invite Code: cptpovgohttps://www.runninghub.ai/user-center/1993809672580788226?inviteCode=cptpovgo
    喜欢我的作品吗?🎨 我会持续免费分享更多优质内容✨,欢迎支持一下,请我喝杯咖啡吧!💖 Like my work? 🎨 I share all my work for free ✨, so support me by buying me a coffee 💖!

    Wanna leave me a tip? Drop by my discord channel for links and special perks. 想赞助一下?来我 Discord,里面有链接和特别福利哦!

    Fanvue: https://www.fanvue.com/catlover1937

    Ko-fi: ko-fi.com/catlover1937

    Please do not redistribute without consent. 如非作者同意禁止搬运。


    2026-07-13

    🎉 Moody Real Mix V8 is Here! ✨

    Hello everyone, welcome to the release of my newest Zimage model — Moody Real Mix V8!

    ...Just kidding! 😂

    But honestly... could you tell these demo images were actually generated with Krea2? 👀

    Once again, I'm amazed by how incredibly flexible the Krea2 base model is. I genuinely think the overall aesthetic has reached the point where it can compete with my own ZIT Real series. On top of that, it delivers surprisingly strong NSFW capabilities, handles difficult poses much more reliably, and rarely falls apart. It honestly feels like the complete package. 🔥

    🎁 Bonus: V2 also has much better support for generating different ethnicities!

    To me, this feels like a milestone release. There's still plenty of room for improvement, of course, but it's exciting to see just how far we've come.

    One thing I've never fully understood is why so many people obsess over tiny pixel-level imperfections. Personally, I believe a model shouldn't be judged by whether you can zoom in 400% and find a flaw.

    Instead, the real question is:

    When you first look at an image, do you instantly think, "Wow, I love this"?

    If the answer is yes, then I think the model has already succeeded. ❤️

    As always, I'll continue actively developing both of my model series:

    Moody Cutie Mix — focused on the trendy "social media girl" aesthetic with beautiful stylized visuals.

    📸 Moody Krea2 Mix — focused on realism while maintaining an attractive overall aesthetic.


    📖 What is this model designed for?

    If you're new to my models, here's the philosophy behind this series.

    This model is designed around the "网感萌妹" (trendy cute social media girl) aesthetic.

    Rather than pursuing strict realism, it embraces the polished, heavily filtered beauty style that's commonly seen across modern social media. 📸💕

    The goal isn't to recreate reality—it's to create images that immediately catch your eye and simply look beautiful.

    While this model is not specifically trained for NSFW generation, it is NSFW unlocked, meaning it retains the original Krea2 NSFW capabilities out of the box without additional censorship.

    🎉 Moody Real Mix V8 正式发布!✨

    大家好,欢迎来到我的全新 Zimage 模型 —— Moody Real Mix V8

    ……骗你的啦!😂

    不过说真的,你们能一眼看出来这些展示图其实都是 Krea2 生成的吗? 👀

    我再次被 Krea2 的可塑性震惊到了。我觉得现在它的整体美感已经完全可以和我的 ZIT Real 系列掰掰手腕了,而且 NSFW 能力依然很强,人物姿势稳定,不容易崩,整体表现真的非常优秀。🔥

    🎁 Bonus:V2 对不同种族人物的支持也提升了不少!

    对我来说,这绝对算得上是一个里程碑式的进步。当然,它还有很多可以继续优化的地方,但看到它成长到今天,我还是非常满意的。

    另外,我一直不是特别理解为什么那么多人喜欢数毛、放大看像素。🤏

    在我看来,一个模型真正的价值,不应该是放大 400% 去找瑕疵,而是:

    当你第一眼看到它生成的图片时,会不会马上觉得:「哇,这张真好看!」

    如果能够做到这一点,我觉得它就是一个成功的模型。❤️

    未来我也会继续坚持双线更新:

    Moody Cutie Mix —— 主打网感、美感和二次元氛围。

    📸 Moody Krea2 Mix —— 主打偏写实风格,同时兼顾优秀的整体美感。


    📖 这个模型到底适合什么?

    如果你是第一次接触我的模型,这里再简单介绍一下它的定位。

    这个系列围绕 「网感萌妹」 美学打造。

    它并不是追求绝对写实,而是追求如今社交媒体上最常见的那种经过精致修饰、滤镜优化后的高级美感。📸💕

    目标不是百分之百还原现实,而是生成第一眼就足够吸引人、让人觉得舒服、漂亮、有氛围感的图片。

    另外,本模型并没有专门针对 NSFW 内容进行训练,但它没有进行 NSFW 限制,因此完整保留了 Krea2 原生的 NSFW 能力,可直接使用。


    🌸 Moody Cutie Mix is Here! 💖

    🇺🇸 English

    Hello everyone! 👋

    I'm excited to introduce my newest Krea2 model: ✨ Moody Cutie Mix ✨

    This model is designed around the "网感萌妹" (trendy cute social media girl) aesthetic. Rather than pursuing realism, it embraces the heavily filtered, polished Asian beauty style that's commonly seen across modern social media. 📸💕

    While this model is not specifically tuned for NSFW generation, it is NSFW unlocked, meaning it still retains the base Krea2 NSFW capabilities out of the box.

    I wanted this model to complement my main Krea2 release rather than replace it. ❤️


    ❓ FAQ

    1️⃣ What's the difference between my main Krea2 model and Moody Cutie Mix?

    Main Krea2 Model

    • 📷 Realism focused

    • 🔓 NSFW unlocked

    • 🔥 NSFW enhanced

    • 👤 Better for realistic portraits and mature photography styles

    Moody Cutie Mix

    • 🌸 Cute social media aesthetic

    • ✨ Heavy Asian beauty filter style

    • 💕 Great for sweet, wholesome, everyday images

    • 📱 Perfect for influencer-style photos, selfies, lifestyle shots, and trendy portraits


    2️⃣ Will you continue updating the main Krea2 model?

    ✅ Yes!

    Both Moody Cutie Mix and my main Krea2 model will continue receiving regular updates and improvements.


    3️⃣ Which model variant should I use?

    • BF16 – Best image quality and maximum compatibility. Requires ~26 GB+ VRAM. Recommended if you have enough VRAM and want the highest possible quality.

    • FP8 – Best all-around choice for RTX 40-series and RTX 50-series. Excellent balance of speed, VRAM usage, image quality, and compatibility.

    • INT8 – Recommended for RTX 30-series. Lower VRAM usage than BF16 while maintaining very good image quality.

    • NVFP4 – Designed for RTX 50-series (Blackwell) GPUs. Lowest VRAM usage and fastest on supported hardware, though some custom nodes may not fully support it yet.

    🚀 Quick Recommendation

    • RTX 30xx → INT8

    • RTX 40xx → FP8

    • RTX 50xx (Blackwell) → NVFP4

    • Maximum quality → BF16 (26 GB+ VRAM)


    4️⃣ Which workflow should I use?

    I highly recommend using this model with my Moody Krea2 Simple Workflow.

    🌱 If you're completely new to ComfyUI, the Moody Krea2 Minimal Workflow is also a fantastic place to start.


    5️⃣ What is the MXFP8 version?

    It's actually INT8 ConvRot.

    Civitai currently doesn't provide INT8 as a selectable model format, so MXFP8 was chosen simply as the closest available category.


    6️⃣ Does this model have any racial bias?

    By default, the model tends to generate Asian-looking subjects if no ethnicity is specified.

    However, when you explicitly specify an ethnicity or race in your prompt, it generally follows the prompt accurately.


    ⚠️ Known Issues

    These will be improved in future updates.

    • ✋ Hands and feet can sometimes look awkward.

    • 🎭 Skin may occasionally appear as a strange mix of ultra-smooth beauty-filter skin and natural skin imperfections. I suspect my LoRA and the Krea2 base model are still "fighting" each other a bit here 😅.


    ❤️ Final Words

    Thank you so much for all of your support!

    I genuinely hope you enjoy this model and have lots of fun creating with it.

    Happy generating! 🌸✨
    Peace! ✌️❤️


    🌸 Moody Cutie Mix 发布啦!💖

    🇨🇳 中文

    大家好!👋

    今天给大家带来我的全新 Krea2 模型:

    ✨ Moody Cutie Mix ✨

    这是一款主打 「网感萌妹」 风格的模型。

    它并不是以真实摄影为目标,而是更加偏向于当下社交媒体流行的亚洲美颜滤镜风格,整体更加甜美、精致、有少女感。📸💕

    虽然本模型并没有专门针对 NSFW 内容进行强化训练,但已经解锁 NSFW,因此依然保留了 Krea2 原本的 NSFW 能力。

    我希望它能与我的主力 Krea2 模型形成互补,而不是互相取代。❤️


    ❓ 常见问题(FAQ)

    1️⃣ Moody Cutie Mix 和主力 Krea2 模型有什么区别?

    主力 Krea2 模型

    • 📷 偏真实摄影风格

    • 🔓 已解锁 NSFW

    • 🔥 对 NSFW 有进一步增强

    • 👤 更适合真实人像与摄影作品

    Moody Cutie Mix

    • 🌸 主打网感萌妹风格

    • ✨ 偏亚洲美颜滤镜审美

    • 💕 更适合甜美、日常、治愈系图片

    • 📱 非常适合自拍、社交媒体、美少女、生活照等风格


    2️⃣ 主力 Krea2 模型还会继续更新吗?

    当然会!✅

    Moody Cutie Mix主力 Krea2 模型 都会持续更新与维护。


    3️⃣ 我应该下载哪个模型版本?

    🇨🇳 推荐版本

    • BF16 – 画质最佳、兼容性最好。需要约 26 GB 以上显存。 如果显存充足,推荐使用。

    • FP8RTX 40 系列和 RTX 50 系列最佳全能选择,在速度、画质、显存占用和兼容性之间表现最均衡。

    • INT8 – 推荐 RTX 30 系列 用户使用,显存占用更低,同时仍能保持非常不错的画质。

    • NVFP4 – 专为 RTX 50 系列(Blackwell) 优化,占用显存最低、速度最快,但部分自定义节点目前可能尚未完全兼容。

    🚀 快速推荐

    • RTX 30xx → INT8

    • RTX 40xx → FP8

    • RTX 50xx → NVFP4

    • 追求最高画质 → BF16(26 GB+ 显存)


    4️⃣ 推荐使用哪个工作流?

    我最推荐搭配 Moody Krea2 Simple Workflow 使用。

    🌱 如果你是新手,也可以直接使用 Moody Krea2 Minimal Workflow,简单易上手,非常适合入门。


    5️⃣ 什么是 MXFP8 版本?

    其实它就是 INT8 ConvRot

    由于 Civitai 没有提供 INT8 这个模型格式选项,所以我只能选择最接近的 MXFP8 分类来发布。


    6️⃣ 模型会有种族偏向吗?

    如果提示词没有指定种族,模型通常会默认生成偏亚洲脸。

    不过只要你明确指定种族或民族,模型通常都能较准确地按照提示词进行生成。


    ⚠️ 已知问题

    后续版本会继续优化。

    • ✋ 手和脚偶尔会画得不太自然。

    • 🎭 有时候皮肤会出现一种很奇怪的效果:一部分像开了重度美颜,一部分又保留了真实皮肤纹理。我猜应该是我的 LoRA 和 Krea2 Base Model 还在"互相较劲" 😅。


    ❤️ 最后想说的话

    感谢大家一直以来的支持!🙏

    希望大家喜欢这款模型,也希望它能给大家带来更多创作灵感。

    玩得开心,我们下个版本见!🌸✨

    Peace!✌️❤️

    Description

    FAQ

    Comments (18)

    HcupManiaJul 14, 2026· 4 reactions
    CivitAI

    您好。我是韩国用户,这段反馈是通过翻译工具整理的,如果有表达不准确的地方,还请谅解。

    我经常使用您制作的 Moody 系列模型,也很喜欢这些模型,所以想反馈一个可能与 INT8 ConvRot 版本兼容性有关的问题。

    测试模型:

    moodyCutieMixKrea2_v20INT8CONVROT.safetensors

    测试环境:

    Forge Neo

    Python 3.13.12

    PyTorch 2.12.0 + ROCm 7.15

    Radeon RX 7900 XTX 24GB

    Qwen3-VL 4B FP8 文本编码器

    Qwen Image VAE

    主要问题是,Forge Neo 似乎没有把这个模型识别为原生 INT8 或 FP8 模型。

    Moody 模型加载时的日志如下:

    Using MixedPrecision for Model
    Diffusion Model: {storage: torch.bfloat16, computation: Mixed}

    即使我手动把 UNet storage dtype 设置为 torch.float8_e4m3fn,模型最终仍然以 BF16 / MixedPrecision 方式加载。

    作为对比,另一个 Krea2 模型可以正常识别为:

    Using Detected Model Data Type: torch.float8_e4m3fn
    Diffusion Model: {storage: torch.float8_e4m3fn, computation: torch.bfloat16}

    LoRA 的运行情况也有明显差异。

    在 Moody 模型中,日志显示:

    with on_the_fly = False

    不使用 LoRA 时,大约是每步 9.8 秒。
    加载 LoRA 后,速度会下降到大约每步 70 秒,而且 LoRA 的效果也比较弱或不稳定。

    在对比模型中,日志显示:

    with on_the_fly = True

    相同的 LoRA 可以正常加载全部 256 个 key,并以大约每步 4.9 秒的速度运行。

    我也关闭了强制 VRAM offload 相关选项重新测试,但问题仍然存在。因此,这个问题似乎不只是显存或 offload 设置造成的。

    我猜测这个 Moody INT8 ConvRot 模型可能使用了较旧或不同的量化格式,或者其中的 scale、ConvRot 等量化 metadata 与目前 Forge Neo / ComfyUI 的原生 INT8、FP8 加载方式不完全兼容。

    模型本身仍然可以生成图片,所以文件应该不是完全损坏。但是它似乎会回退到速度较慢的 MixedPrecision 路径,特别是在加载 LoRA 后问题非常明显。

    方便的话,能否检查一下这个 INT8 ConvRot 版本使用的转换方式,以及它是否与最新的 ComfyUI / Forge 原生量化格式兼容?

    因为我一直很喜欢并经常使用您的 Moody 模型,也希望之后可以继续搭配 LoRA 使用,所以特意向您反馈这个问题。

    谢谢。

    catlover1937
    Author
    Jul 14, 2026· 1 reaction

    这个问题在Moody Krea2 Mix也谈到过,这是一个比较头疼的问题。我如果按照正常的bf16 转换成fp8跟int8,comfyui使用画质会变得非常塑料。但是如果我用comfyui的基础模型来合成,你们这些forge的用户又识别不了。

    我能想到的临时折中方法就是再出一版专门为forge 定制的量化,但是我不知道forge用户到底有多少,值不值得这么做

    archlliesJul 14, 2026
    CivitAI

    請問v2會有BF16的版本嗎?

    catlover1937
    Author
    Jul 14, 2026

    有的,文件损坏了 正在重新做 还得下载本地然后再上传

    archlliesJul 14, 2026

    @catlover1937 感謝~辛苦了

    EibbaNewJul 14, 2026
    CivitAI

    感谢萌妹王的辛勤劳作,提交一个反馈,V2搭配minimalV3工作流,有时候会出现奇怪的图案,大概10张会出现一两次。

    catlover1937
    Author
    Jul 14, 2026· 1 reaction

    对啊 不知道是啥原因 还得继续琢磨

    EibbaNewJul 14, 2026

    @catlover1937 另外问下krea2版的V4对比V3清晰度和锐度都提高了很多,是换了新的方法吗?能用到zit那边么?

    catlover1937
    Author
    Jul 14, 2026

    @EibbaNew zit已经很清晰了呀

    EibbaNewJul 14, 2026

    @catlover1937 可能是我的错觉吧,总感觉清晰度不如krea2,眼睫毛,皮肤细节有点糊

    xiwsgg232Jul 14, 2026

    @catlover1937 我也出现了很多次这个现象,我还以为是我的显卡显存不足爆显存造成的呢

    catlover1937
    Author
    Jul 14, 2026

    @xiwsgg232 不好说不好说 换一个种子就能解决,我感觉可能得重新筛选一下训练集

    10905081363Jul 14, 2026
    CivitAI

    大佬还是强的,我记得在krea2主模型v1还是v2版本提过,说这模型头皮发根那里很奇怪、很假,佬应该是加了增加发量的lora,V4就已经解决了,这个萌妹版v1应该是权重多了点导致发型偏多、脏、乱,看v2例图也解决了这个问题-3-

    VicePaletteJul 14, 2026
    CivitAI

    I'll give it a spin. Any prompt ideas to get out most of it?

    ILoveAsianGirlsJul 14, 2026
    CivitAI

    Way better than beta :) GJ :)
    But.. but... also I like zooming 800% and seeing my creations look like real stuff :P It is even more cute when I have cute picture and I can zoom in on face and it looks like different made face picture, because there are so much pixels it still looks good :D

    NewarkianJul 16, 2026
    CivitAI

    Which Moody checkpoint is the best? among all the krea 2.

    catlover1937
    Author
    Jul 16, 2026· 1 reaction

    I think Krea2 mix V4 is the safest bet, Cutie Mix V2 has its perks but it's not general enough for a daily use model. Depends on your need also

    NewarkianJul 17, 2026

    @catlover1937 Thank you.

    Checkpoint
    Krea 2

    Details

    Downloads
    23
    Platform
    CivitAI
    Platform Status
    Available
    Created
    7/14/2026
    Updated
    7/21/2026
    Deleted
    -

    Files

    moodyCutieMixKrea2_v20BF16.safetensors

    Mirrors