Drop in my Telegram Group and Discord group for chats, questions, feedbacks. 欢迎加入纸飞机和Discord群组。

RH体验地址（邀请码Invite Code: cptpovgo）：https://www.runninghub.ai/user-center/1993809672580788226?inviteCode=cptpovgo
喜欢我的作品吗？🎨 我会持续分享更多优质内容✨，欢迎支持一下，请我喝杯咖啡吧！💖 Like my work? 🎨 support me by buying me a coffee ☕️! Drop by my discord channel for supporter perks and download links. 想赞助一下？来我 Discord，里面有链接和特别福利和专属付费模型下载哦！

CivitAI 邀请码 Referral Code: R44VYP99 Extra 2.5k Buzz on membership purchase，买会员送2千5闪电!

Ko-fi: ko-fi.com/catlover1937

Please do not redistribute without consent. 如非作者同意禁止搬运到其他平台，如国内Up主需要转发网盘请备注出处。

WOULD YOU PLEASE CLICK THE LIKE BUTTON? 拜托动动您可爱的小手点个赞好吗？

2026-11-04

MoodyMossMix V4.1 🎨

Back with another update! For V4.1, the focus was stripping out redundant weights and fixing the skin over-baking while keeping that signature V3 mood and dynamic lighting intact.

Huge props to the MOSS AI team—seriously, the depth of visual knowledge packed into this ~5B parameter base model is mind-blowing. Because MOSS already understands so many concepts natively, I was able to prune out a ton of legacy data previously prepped for Krea 2. The result? A cleaner checkpoint, less clutter, and noticeably better prompt adherence.

🛠️ What's Changed in V4.1:

Goodbye Plastic Skin: Dialed back the heavy "oil-slick" sheen and over-baked specular highlights. Skin tones and textures look significantly more natural and cinematic now.
Pruned & Cleaned Weights: Stripped out redundant training concepts that MOSS already covers. Cleaner latent space, better prompt compliance.
V3 Vibe Preserved: Kept the dark, moody contrast and atmosphere you guys liked from V3, with noticeably more flexibility across different prompts and styles.

🛡️ Why Release BF16? (And the Truth About Model Merging)

Let’s be real for a second: a lot of popular creators refuse to release full-precision or unquantized checkpoints because they’re terrified of people scraping, ripping, or merging their weights into "new" models without credit.

I’ve never believed in holding back the community like that. Even back during Krea 2, I was among the very first creators to drop the full BF16 version. I want you to have the real weights in your hands.

Speaking of which—if you want a deep dive into the weird reality of what’s happening in modern diffusion workflows, check out my new article: "How Models Are Merging Themselves" (link). It breaks down how the architecture and ecosystem are evolving beyond manual merges.

📦 Downloads & Fair Access (Read Before Buying!)

I see the comments about having to buy again every time an update drops, and I completely get the frustration. Civitai doesn't have an option to let you "pay once and get all future updates forever," so here is how I'm handling it so no one gets burned:

FP8 & INT8 (100% Free): The quantized versions are free for everyone. Honestly, for 95% of daily use and home setups, these run faster, save precious VRAM, and the quality difference is barely noticeable.
Already bought V3.3 BF16? Do not pay again! I’ve uploaded the new V4.1 file directly under the V3.3 BF16 download section, so you can just grab the update straight from your previous purchase.
Support the Craft: If you want all my private checkpoints, experimental merges, early builds, and future releases in one place without buying individual drops on Civitai, consider becoming a supporter directly.

Drop your renders and feedback below—excited to see what you guys create with 4.1! 🚀

MoodyMossMix V4.1 🎨

新版本搞定，正式发布！这次 V4.1 的核心重点就是：清理冗余权重 + 爆改油腻感，同时完整保留大家最爱的 V3 情绪光影与暗调质感。

先给 MOSS AI 团队 递根烟，致敬一波。说实话，作为一个才 5B 参数量左右的模型，底模里沉淀的先验知识量简直离谱。正因为 MOSS 原生底模对概念的理解已经足够扎实，我彻底清理掉了之前为了 Krea 2 特别塞进去的一大堆训练概念。瘦身之后，模型不拖泥带水，提示词响应和泛化性直接上了一个台阶。

🛠️ V4.1 核心更新点

告别“猪油脸”： 专门针对皮肤高光过曝、油光满面的问题做了深度修复。现在的皮肤质感更加自然通透，不再有那种重度后期、过度烘焙（over-baked）的塑料假人感。
精简权重与概念： 彻底移除了大量 Krea 2 留下的冗余 tokens。底模懂的就不重复塞，隐空间更干净，画风更听话。
继承 V3 核心美学： 依旧是熟悉的冷峻、氛围感和电影级调色，但在不同风格间的扩展自由度更高，不再容易锁死死板画风。

🛡️ 关于坚持发 BF16，以及“套壳缝合”这件事

大家心知肚明，圈子里不少大作者压根不舍得放全精度甚至是高质量量化版，说白了就是怕别人偷权重、拿去随便 merge 几下就当成自己的“原创”模型发出来。

但玩社区没必要这么抠搜。早在 Krea 2 那会儿，我就是最早公开分享 BF16 版本的作者之一。好东西就该让大家拿到满血体验。

顺便安利一下，如果你对现在模型融合的生态好奇，强烈建议去读读我刚写的新文章：《模型是如何自我融合的》（How Models Are Merging Themselves） ，带你扒一扒目前各种模型架构演进的底层真相。

📦 下载与老用户避坑指南（下单前先看！）

看到评论区总有兄弟吐槽：“刚买了旧版，新版就出了，难道每次更新都要重新交一次钱？”

我完全理解这有多蛋疼。但 Civitai 目前平台机制就那样，不支持“一次性买断永久订阅”。所以我直接按自己的方式给大家搞福利，绝不割老粉：

FP8 和 INT8 全员免费白嫖： 讲良心话，对于 95% 的朋友日常出图、或者中低显存跑图来说，这两个量化版不管是速度还是显存占用都香得多，肉眼画质基本看不出差距，直接下就行，不用花一分钱。
已经买过 V3.3 BF16 的兄弟别重复花钱！ 我已经把这次的 V4.1 模型文件直接挂到了 V3.3 BF16 的下载列表里，之前买过的直接去原购买处重新下载即可，老用户无缝白拿。
想一次性包圆所有资源？ 如果你不想每次都在平台跟着零散更新折腾，或者想提前玩到我所有的私藏底模、实验性融合版本和内测模型，欢迎直接去支持我的赞助计划，感谢每一位给算力回血的真爱粉！

快去跑图吧，非常期待在返图区看到大家的全新作品，记得多发图片和反馈！🚀

2026-10-28

🚀 Moody MOSS Mix V3: Rebuilt, Rebalanced & De-Oiled! 🎨✨

Hey everyone! 👋 After weeks of pushing my potato PC to its absolute thermal limit 🥔🔥, the training is finally 100% done, and Moody MOSS Mix V3 is officially here!

MOSS has been out in the wild for a while now, and I’ve finally dialed in the core foundational LoRAs to hit the visual standard I was looking for, V1 was about 30% and V2 was about 60%-70%.

🤔 The Story So Far: Why Did V1 & V2 Turn Everyone Asian?

In V1 and V2, I reused my training set from Krea 2. The catch? Krea 2 had a strong default Western bias, so I originally pumped it full of Asian facial aesthetics to balance the scales.

However, transferring that exact dataset over to MOSS created an extreme race bias in the opposite direction. 😅 My DMs and comments were flooded with: "Hey, why does everyone look Asian now?" You were totally right. I heard you loud and clear, and V3 should've improved from this bias quite a bit.

🛠️ What’s New in V3 (The 4 Big Updates)

⚡ Dual Turbo LoRA: Upgraded from a single Turbo LoRA to a Dual Turbo LoRA setup. You still get faster 25-step generation speeds, but with a massive upgrade: that plastic, greasy/oily AI skin texture is practically gone! The custom LoRA blend is now finely tuned specifically for 25 steps.
🌍 Completely Rebuilt Diversity Dataset: Reworked the entire facial aesthetic dataset from scratch. The Asian ratio has been drastically reduced, introducing a truly balanced, diverse mix of ethnicities and features.
🧬 Base Model Blend (Anti-SFT Bias): Blended a portion of the original base model back in. As many of you know, SFT naturally carries a slight Asian portrait drift—this blend effectively neutralizes that bias.
✂️ Stripped the Heavy "Moody" Signature Filter: In Moody Real Mix, I loved using that signature high-contrast, ultra-vibrant, cinematic shadow filter from the Zimage era. But feedback showed a lot of you actually prefer raw, unfiltered realism over heavy stylization. (Honestly, I was surprised—life is already rough enough, why wouldn't you want extra perfections in AI art?! 😂) But your feedback runs the show: the heavy filter is gone, leaving clean, authentic aesthetics. (substitued with my youthful lense filter at a very low weight)

⚙️ Recommended Settings

Steps: 25 (Optimized for this speed)
CFG Scale: 3.5
Sampler: Euler Simple

💬 Final Thoughts

At the end of the day, community feedback makes these models better. If you wanted realism over hyper-stylization, you got it! I genuinely hope you enjoy V3, and hopefully, the results make up for my skyrocketing electricity bill and countless hours of dataset curation. ⚡💸

Drop your generations, reviews, and gives a thumb up below! Happy generating! 🥂🎉

🚀 Moody MOSS Mix V3 正式发布：去油腻、真多元、回归真实美感！🎨✨

各位小伙伴们好！👋 经过几星期让我家的两台破电脑风扇狂转、直接推到散热极限之后 🥔🔥，几个核心底层 LoRA 终于全部搞定达到100%训练度，Moody MOSS Mix V3 终于能和大家见面了！

MOSS 模型发布也有一阵子了，前两个版本出来后，我收到了很多反馈，今天这个版本算是完整了吧（V1大概 30%训练度， V2大概是60-70%）。

🤔 关于前作的“血统”问题：为什么之前生图全偏亚洲面孔？

前两个版本里，我偷懒沿用了之前做 Krea 2 时的训练集。熟悉 Krea 2 的朋友都知道，那是一个底层极度偏欧美/西方的模型，所以我当时为了美化亚洲人像，往里面塞了海量的亚洲人素材。

结果这套素材直接搬到 MOSS 上，直接水土不服，产生了严重的 人种偏向（Race Bias） 😅。很多小伙伴私信和评论区都在跟我反馈这个问题，我也完全意识到了。既然发现了问题，V3 就彻底重构！

🛠️ V3 版本的四大核心调整

⚡ 双 Turbo LoRA 架构，告别塑料油腻感： 从原先的单一高权重Turbo LoRA 改为低权重双 Turbo LoRA。在依然保持 25 步快速出图 的前提下，大幅消除了 AI 生图常见的“塑料反光”和“重度油腻感”！同时底层 LoRA 也专门针对 25 步进行了重新调配与适配。
🌍 重构人物美感数据集： 全新制作了人像数据集，大幅削减了亚洲人像素材的极端占比，融入了更多元化、更具质感的世界人种素材。
🧬 融入原生 Base 模型，对冲 SFT 偏移： 融入了一定比例的原版 Base 模型。之前就和大家聊过，SFT 模式本身就自带一些亚洲人偏向，融入 Base 之后能有效中和这种跑偏现象。由此还增加了不少真实感。
✂️ 果断移除强风格滤镜 LoRA： 很多用过我 Moody Real Mix 的老粉应该知道，我一直很偏爱那种招牌式的 Moody Zimage 风格——色彩高饱和、阴影强对比、自带浓郁电影滤镜。但我发现现在越来越多的朋友并不想吃这一套，大家更想要的是原汁原味的真实感与写实度，反而不太喜欢过度美化的滤镜感。(我当时还挺诧异的：生活都已经这么破旧不堪了，大家在 AI 世界里居然不想追求点极致美好吗？！😂) 不过吐槽归吐槽，用户的真实体验才是最重要的。既然大家想要真实，那我就果断删掉风格滤镜，专注打磨底模的高级美感！（换成了少量的清新滤镜）

⚙️ 推荐生成参数

生成步数 (Steps): 25 (本版本专属黄金步数)
提示词引导系数 (CFG Scale): 3.5 - 5.0
采样器 (Sampler): Euler Simple

💬 创作者碎碎念

模型好不好，归根结底大家用得顺手才是王道。你们想要的真实、清爽、不偏色，我都尽最大努力做进 V3 里了。

希望新版本大家能够喜欢，也希望这次 MOSS 的表现，能对得起我电脑没日没夜烧掉的高昂电费和通宵洗数据集的精力吧！⚡💸

欢迎大家在评论区狂甩返图、提建议、点赞👍！祝大家玩得开心，把把出神图！🥂🎉

Quick note on "Asian Bias":

It's mainly due to three reasons:

1. Training dataset bias. My training dataset leans toward Asian faces, and with smaller 5B models like Moss, that bias seems to get amplified. It reminds me of when I was making my Z-Image models. It's surprisingly hard to find datasets with the specific aesthetic I'm looking for that feature Western faces, if you know what I mean.

2. Certain keywords reinforce the bias. Words like "cute," "cosplay," "black hair," "big eyes," etc., seem to push the model even further toward Asian features. From my testing, the base model itself already has quite a strong Asian bias.

Try this prompt with the base SFT model and see what you get: A cute cosplay Russian girl with black hair.

3. The Turbo LoRA makes it even more pronounced. The Turbo LoRA I've found to work best is likely trained on a dataset that also leans heavily toward Asian faces, which further exaggerates the bias.

As for your question: Yes, I believe this can be improved in future versions. In the meantime, you can also use this Lora from @iamddtla with negative weight https://civarchive.com/models/3016507/pornmaster-moss-asian-slider to help mitigate the bias.

2026-10-23

🎉 Moody MOSS V2 Mix Release!

After countless hours of training, testing, tweaking, and probably way too much frustration, Moody MOSS V2 Mix is finally here! 🚀

I think I've finally managed to transfer a pretty significant portion of my base resources over to MOSS, and for now, I've decided to temporarily pause further training. I still have some promised Krea 2 projects to finish before the end of the month, so it's time to shift my focus back to those.

Personally, I'm quite satisfied with how V2 turned out, and honestly, this will likely be my best MOSS release for a while, at least until the official Turbo model arrives or we start seeing some really good Turbo LoRAs.

🧪 The Struggle Behind V2

One of the biggest challenges with MOSS right now is its painfully slow generation speed.

And with my rather crappy testing setup, fine-tuning this model has been an absolute test of patience. 😂

Imagine making one tiny adjustment, waiting nearly 2 minutes just to see the result, realizing something still isn't quite right, and doing it all over again.

Now imagine repeating that process over a thousand times.

Yeah... that's pretty much been my life lately. 💀

I've spent an enormous amount of time trying to inject the signature Moody style into MOSS, and I have to admit, this was much harder than I originally expected.

Here's something interesting I've learned about MOSS:

This model is basically a sponge. 🧽

It absorbs concepts, visual characteristics, and styles incredibly quickly. In fact, its ability to pick up new information is genuinely impressive.

But that strength is also one of its biggest challenges.

It's ridiculously easy to overcook. 🔥

Push the training just a little too far, and things can start going sideways. Finding that sweet spot between strong style transfer, flexibility, and preserving the model's original capabilities has been surprisingly difficult.

Compared to many of the models I've worked with in the past, MOSS has honestly been one of the trickiest models to fine-tune properly.

It's been a lot of trial and error, a lot of waiting, and a lot of questioning my life choices. 😂

❤️ What's Next?

For now, I'm happy with where Moody MOSS V2 Mix has landed.

I still think there's plenty of untapped potential in MOSS, and I'm particularly excited to see what becomes possible once faster generation options are available. But until then, I think I've pushed this version about as far as I'm willing to go with my current setup.

A huge thank you to everyone who's been following my releases, testing my models, and supporting my work. Your support genuinely keeps me motivated to continue experimenting and improving.

I hope you enjoy Moody MOSS V2 Mix as much as I enjoyed creating it (minus the suffering, of course). 😂❤️

Have fun creating, and I'll see you in the next release!

— Moody ✨

🎉 Moody MOSS V2 Mix 正式发布！

经过无数个小时的训练、测试、调整，以及数不清的崩溃瞬间，Moody MOSS V2 Mix 终于来了！ 🚀

我觉得目前已经成功将相当一部分原有的基础训练资源迁移到了 MOSS 上，所以决定暂时告一段落，暂停进一步的训练。

毕竟这个月底之前，我还有一些之前答应大家的 Krea 2 项目需要完成，是时候先把精力转回去了。

就个人而言，我对 V2 的最终效果还是相当满意的。

而且，这很可能会是我近期能拿出的最好的 MOSS 版本了，至少在官方 Turbo 模型正式发布，或者出现一些足够优秀的 Turbo LoRA 之前，我应该不会再进行大规模的更新。

🧪 V2 背后的折磨

目前 MOSS 最大的问题之一，就是它实在太慢了。

尤其是配合我那套不怎么样的测试设备，整个微调过程简直就是一场耐心极限挑战。😂

想象一下：

你只是对模型做了一点点微小的调整，然后就得等上将近 2 分钟，才能看到生成结果。

看完之后发现效果还是不太对，于是继续修改，再等两分钟。

然后把这个过程重复 一千多次。

没错，这差不多就是我最近的日常。💀

为了尽可能将我标志性的 Moody 风格 融入 MOSS，我确实花费了大量的时间和精力。

但不得不承认，这个过程远比我最初预想的困难。

在训练过程中，我发现 MOSS 有一个非常有意思的特点：

这个模型简直就像一块海绵。 🧽

它吸收新概念、视觉特征和风格的速度快得惊人。说实话，它的学习能力确实让我印象深刻。

但问题就在于，学得太快，有时候也不是什么好事。

MOSS 真的太容易被练过头了！ 🔥

稍微多训练一点，效果就可能开始跑偏。

如何在强化 Moody 风格的同时，保留模型原本的灵活性、泛化能力以及对提示词的理解能力，成了整个训练过程中最难把握的部分。

相比我过去制作过的许多模型，MOSS 可以说是我遇到过最难驾驭的微调模型之一。

整个过程充满了反复试错、漫长等待，以及无数次对人生选择的怀疑。😂

❤️ 接下来呢？

至少目前，我对 Moody MOSS V2 Mix 的表现已经比较满意了。

我依然认为 MOSS 有着非常大的潜力，还有许多值得继续探索的空间。尤其是未来如果能够拥有更成熟、更快速的生成方案，我相信它还能带来更多惊喜。

但在那之前，以我目前的设备和时间成本来说，我觉得已经把这个版本打磨到了一个自己能够接受的极限。

最后，真的非常感谢所有一直关注我的更新、测试我的模型，以及支持我创作的朋友们。

你们的支持一直都是我持续折腾这些模型的重要动力。

希望大家能够喜欢这次的 Moody MOSS V2 Mix，也希望你们能享受到它带来的创作乐趣——当然，最好不要体验我训练它时经历的那些痛苦。 😂❤️

祝大家玩得开心，我们下次更新再见！

— Moody ✨

2026-10-20

I'm sorry I haven't shown you guys the true potential of my Moody models until now! 😭❤️

With this absolutely stunning 4K HD Workflow, the results are on a whole different level of realism! The details, the textures, the overall quality... everything looks AMAZING!

I'm genuinely blown away by the results and absolutely in love with it! 😍🔥

Please, please go check it out! You won't regret it! ❤

真的很抱歉，一直以来都没能让大家见识到 Moody 模型真正的实力！ 😭❤️

这次全新的 4K HD 高清工作流，真的把写实效果提升到了一个全新的高度！无论是皮肤质感、细节表现，还是整体画质，都惊艳到我了！

说实话，连我自己都被生成效果震撼到了，简直爱不释手！😍🔥

大家一定一定要去试试看！真的不会让你们失望！ ❤️

https://civarchive.com/models/3012555/moody-moss-4khd-workflow?modelVersionId=3419146

2026-10-18

🐱💨 Moody Moss Mix V1 — 25 Steps Edition 🎨✨

🚀 The 25 Steps Version is here!

I'm excited to release Moody Moss Mix V1 (25 Steps)! 🎉

Unlike the previous Turbo version, this version is specifically tuned to perform at 25 steps, delivering significantly improved image quality, richer details, and better overall visual results. 🖼️✨

If you've tried the Turbo version, you'll definitely want to give this one a shot! 🔥 The extra steps really make a difference in graphical quality.

⚠️ Important: Make sure you use it with my paired 25 Steps Workflow for the intended results. The workflow is an essential part of the experience!

Have fun experimenting, and enjoy the new version! 💜🐾

🐱💨 Moody Moss Mix V1 — 25 步版本 🎨✨

🚀 25 步版本正式发布！

很高兴为大家带来 Moody Moss Mix V1（25 Steps）！🎉

与之前的 Turbo 版本不同，这次的版本专门针对 25 步采样进行了调校，能够显著提升画面质量、细节表现以及整体视觉效果。🖼️✨

如果你之前体验过 Turbo 版本，这次绝对值得再试试！🔥 增加采样步数后，画面质感确实会有明显提升。

⚠️ 重要提示： 请务必搭配我提供的**专用 25 步工作流（25 Steps Workflow）**使用，以获得预期效果。工作流是这次体验的重要组成部分！

祝大家玩得开心，尽情探索新版本！💜🐾

(今天太晚了88，too late today, cya) 给你们看看无加速的效果 Preview for Non-turbo Model :

2026-10-17

🌿 Moody Moss Mix V1 Alpha (TURBO) — Preview Release!

Today, I'm bringing you the latest Moody Moss Mix V1 Alpha (TURBO)!

After two nights of training, the base LoRA is roughly 50% through its training progress. I'm releasing an early preview first so everyone can try it out and get a feel for the model's strengths, style, and blending behavior.

🚀 About the Turbo Acceleration

I've experimentally incorporated an acceleration LoRA into this version, allowing it to generate images in approximately 12 steps. However, to be honest, this acceleration LoRA does come with some image quality trade-offs. Skin textures can look overly oily, among other issues.

I'll be releasing a version without the acceleration LoRA later, which should help improve these issues and produce more natural-looking results.

🛠️ Workflows Included

I'll also be providing two workflows for everyone to experiment with. They'll be released as I finish preparing them, so if you don't see them yet, it simply means I haven't finished everything! 😂 Thanks for your patience.

moss minimal workflow

🧪 My Current Impressions

Training LoRAs for this model feels a bit like giving a sponge water: it absorbs information incredibly quickly, but it's also very easy to overfill it until everything starts spilling out.

The difficulty of merging and balancing this model is honestly comparable to what I experienced back when working with ZImage Turbo.

That said, Moss already understands a huge range of concepts and has no content restrictions. This means I don't need to teach it everything from scratch. My main focus is strengthening its understanding of specific body features and refining its overall artistic style.

Of course, we're only on day two or three of exploring this model, so there's still plenty to investigate and improve.

For now, enjoy this early preview! I'll continue releasing updates and refining the model step by step. Stay tuned! 🐱✨

Special thanks to @salttaro .

🌿 Moody Moss Mix V1 Alpha (TURBO) — 预览版发布！

今天给大家带来最新的 Moody Moss Mix V1 Alpha（TURBO）！

经过两个晚上的训练，目前基础 LoRA 大概完成了 50% 左右的训练进度。这次先放出一个预览版本，大家可以先玩玩看，感受一下模型的表现、风格和融合深度。

🚀 关于 Turbo 加速

这次我尝试性地融入了一个加速 LoRA，目前大约 12 步就能出图。不过说实话，这个加速 LoRA 对画质还是有一些负面影响，比如皮肤质感会显得比较油腻。

所以，晚些时候我还会发布一个不含加速 LoRA 的版本，希望能改善这些问题，让画面质感更加自然。

🛠️ 关于工作流

这次也会附带两个工作流，方便大家体验和测试。工作流会陆续放出，如果你暂时还没看到，那就是我还没忙完 😂 请耐心等待！

moss minimal workflow

🧪 目前的训练感受

训练这个模型的 LoRA，有点像给海绵吸水——吸收得特别快，但也很容易吸得太满，甚至溢出来。

在融合难度上，它让我想起了当初折腾 ZImage Turbo 的时候，确实有得一拼。

不过，Moss 本身已经理解了大量概念，而且没有内容限制。因此，我目前不需要从头教它太多东西，主要目标是进一步强化某些部位的理解能力，并调整和完善整体画风。

当然，这还只是第二、第三天的探索阶段，后面还有很多东西值得研究和尝试。

先把这个预览版放出来给大家玩玩！我会慢慢更新、逐步完善，敬请期待。🐱✨

特别感谢芋头哥 @salttaro 的技术支持

Files with SHA256: cfd61b80084728be97424398a2c2aa9a164a1074c151b8cd60af0172c88cfef0