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。
2026-07-10
✨ V4.0 Model Update
A short update note but lots of small stuff! This version includes a newly trained LoRA along with many quality improvements.
🔹 What's Changed
🧠 Integrated a newly trained LoRA for better overall generation quality.
✨ Reduced unwanted skin texture imperfections while preserving natural-looking fine details.
🚿 Reduced the chance of random body fluid artifacts appearing in NSFW generations.
🎨 Improved compatibility and support for external LoRAs.
🚀 For the best experience, pair this model with the new Krea2 Simple Workflow V3 or Minimal Workflow V3. Together they deliver noticeably cleaner outputs, better clarity, and improved overall image quality.
Fixed the FP8 conversion quality degrade issue
Fixed blurry wormlike background issue
✨ V4.0 模型更新
短简介大更新！本次版本加入了全新训练的 LoRA，并带来多项画质优化。
🔹 更新内容
🧠 整合了全新训练的 LoRA，进一步提升整体生成质量。
✨ 在保留丰富细节的同时，减少皮肤纹理瑕疵，让肤质更加自然干净。
🚿 降低 NSFW 场景中身体随机出现体液残留伪影的概率。
🎨 提升对外部 LoRA 的兼容性与表现。
🚀 强烈建议搭配全新的 Krea2 Simple Workflow V3 或 Minimal Workflow V3 使用，可显著提升画面清晰度、干净度以及整体生成质量。
修复FP8转换降画质问题
修复背景失真问题
2026-07-07
🚀 I also added the INT8 CONVROT version! In theory, it should deliver quality similar to Q8 GGUF while running at speeds close to INT8. ⚡
🚀 我新增了 INT8 CONVROT 版本！理论上，它能够提供接近 Q8 GGUF 的画质，同时拥有接近 INT8 的运行速度。⚡
好像很多人不知道每个版本代表着什么，It seems like many of you are unsure of what the models variants mean so here's my take for it:
I was able to rent a cloud server and done a special merge on BF16. Supporters can download version BF16 of this model directly from my discord link. DM me for discord permission tags. 我成功租到了一台云端服务器，并完成了这个模型的 BF16 特别合并版本。支持者现在可以通过我的 Discord 链接直接下载 BF16 版本。如需获取 Discord 下载权限，请私信我领取对应的权限组（Tag）。
FAQ:
Where is int8?
The mxfp8 is the int8, because of civitAI doesn’t have an int8 model type in the dropdown.
What are the special editions (models without Version tag)?
TLDR: download official editions (ones with version tags) for overall better experience.
The Special Edition are heavily inspired by my Asian girl character and style LoRAs. It was never originally intended for release—I created it purely for own enjoyment. However, after sharing some sample results in the community, quite a few people asked if I could release it, so I decided to publish it for those who are interested. It’s designed for a niche audience that enjoys the 网感 (internet aesthetic) beauty filter look.
Hope that clears things up!
2026-07-06
TDLR: V3 is so powerful that i was impressed when testing it. You can safely delete the previous versions now.
✨ Moody Krea2 V3 Release ✨
Hey everyone! 👋
V3 is here, bringing another solid upgrade over V2 with a stronger focus on realism while keeping the signature Moody look. 📸
🔥 What's Improved
🧑🎨 Increased overall realism and more natural-looking results.
🌄 Fixed the lazy background rendering issue from V2. Many of you reported the "worm-like" texture that sometimes appeared in blurred backgrounds—this has been significantly reduced.
💃 Improved human anatomy and more natural poses.
📝 Retrained several LoRAs using both Chinese 🇨🇳 and English 🇬🇧 captions, giving the model a better understanding of prompts and noticeably improving prompt adherence.
✨ Enhanced skin texture for results that look less AI-generated and more photographic.
⚙️ Workflow Update
I've also made a few adjustments to my recommended workflow to better support the model and get the most out of V3.
👉 Please give the new workflows a try for the best experience!
As always, thank you all for the feedback and bug reports—they've been incredibly helpful in making this release possible. ❤️
Enjoy, and have fun generating! 🚀
✨ Moody Krea2 V3 发布啦！✨
大家好～👋
V3 是基于 V2 的全面升级版本，在保留 Moody 风格的同时，进一步提升了整体真实感。📸
我只能说V3。。。很强。。。
🔥 本次更新内容
🧑🎨 整体画面更加真实、自然。
🌄 修复了 V2 中较常见的背景偷懒渲染问题。不少用户反馈虚化背景偶尔会出现类似「蚯蚓纹」的纹理，现在已经大幅改善。
💃 优化人体结构与姿势生成，让人物更加自然。
📝 重新训练了多组 LoRA，并同时使用 中文 🇨🇳 + 英文 🇬🇧 Caption，模型对 Prompt 的理解能力更好，Prompt Adherence 也有明显提升。
✨ 优化皮肤质感，减少 AI 感，让成图更加接近真实摄影效果。
⚙️ 工作流更新
我也针对模型调整了推荐工作流，让 V3 能发挥更好的效果。
👉 建议大家使用新的 Workflow，以获得最佳生成体验！
感谢大家一直以来的反馈和 Bug 回报，很多改进都是来自社区的建议。❤️
祝大家玩得开心，Happy Generating！🚀
2026-07-04
🎉🌸 Special Edition Drop / 特别版发布 💖
🇺🇸 English
Hello everyone! 👋
Here comes another Special Edition. Say no more. 😏
This release is trained on a heavily Xiaohongshu-inspired dataset, featuring East Asian women with beauty-filter aesthetics and curvier proportions. It's made specifically for a niche audience that enjoys this particular style. 💕
If you're not the target audience, feel free to skip this one and wait for the official V3, which is already in the works and will be the more balanced release for everyone. 🚀
Cheers, have fun, and enjoy! 🍻✨
🇨🇳 中文
大家好！👋
又来一个 特别版 啦，不多说了。😏
这次依旧是一个 小红书风格浓厚 的数据集，主打东亚女生、滤镜感颜值，以及更加丰满的身材表现。💕 这是专门为喜欢这种风格的小众用户准备的版本。
如果这不是你的菜，也完全没关系～可以直接等待正在制作中的 官方 V3，那会是更加均衡、适合大众使用的正式版本。🚀
祝大家玩得开心，玩得尽兴！🍻✨
2026-07-01
🎉🚀 Moody Krea2 V2 Official Release! 🌍✨
The wait is over! 💙 Moody Krea2 V2 is officially here!
The Oriental Edition was just a special edition released exclusively for my lovely Discord & Telegram community ❤️. It was made for a specific audience.
The Global Edition should be considered the official V2, featuring the most balanced finetuning I've achieved so far while preserving the signature Moody quality. 😎✨
🎮 GL HF, and enjoy! ❤️
⚠️ PS: About the MXFP8 version
The MXFP8 upload is my attempt at providing an INT8 version. Since Civitai doesn't currently have an INT8 model option, I selected MXFP8 instead.
I followed all the available conversion steps, but I wasn't able to fully verify that it works in every environment. As a result, I can't guarantee compatibility or functionality, so please use it at your own discretion.
🙏🎉🚀 Moody Krea2 V2 正式发布！🌍✨
终于来了！💙 Moody Krea2 V2 正式版 现已发布！
之前的 萌妹版版（Oriental Edition） 只是专门为 Discord 和 Telegram 群友准备的特别版本 ❤️，面向的是比较特定的受众。
而这次的 Global Edition 才是 真正的 V2 正式版 🌎，也是目前为止我调校得最均衡、最能体现 Moody 风格 的版本。😎✨
🎮 祝大家玩得开心，Enjoy！ ❤️
⚠️ PS：关于 MXFP8 版本
MXFP8 是我尝试制作的 INT8 版本。由于 Civitai 目前没有 INT8 的模型选项，所以只能选择 MXFP8 作为上传类型。
我已经按照目前已知的流程完成了转换，但 没有办法完整验证它是否能在所有环境下正常运行。因此，我无法保证兼容性或功能正常，如果有问题还请自行承担风险，感谢理解！🙏
2026-06-30
🌸 Krea2 V2 – Oriental Edition is Here! | Krea2 V2 东方版正式发布！ 🎉
🇺🇸 English
✨ Krea2 V2: Oriental Edition is here!
This SPECIAL EDITION was actually requested by members of our chat group. I originally made it just for my own entertainment and wasn't planning to release it, but after seeing how many people wanted it, I decided to share it with everyone. ❤️
This version focuses on beautiful East Asian female aesthetics with a strong Instagram and Xiaohongshu (RED) inspired vibe. 📸🌸
🔥 What's New
🤝 Huge kudos to @salttaro ! His Better Pussy & Poses V4 plays a major role in this edition on NSFW.
🎨 Added 2 newly trained Moody Aesthetic LoRAs, ported from ZImage for richer visuals.
🔞 Enhanced NSFW prompt adherence for better consistency and prompt accuracy.
NVFP4 Version.
🚀 A regular Krea2 V2 release is coming soon, so stay tuned!
Thank you everyone for your support! ❤️
🇨🇳 中文
✨ Krea2 V2 东方版 正式发布！
其实，这个版本是应聊天群里不少朋友的要求制作的。原本只是做给自己娱乐使用，并没有打算公开发布，不过看到大家都很期待，所以决定分享给大家。❤️
本版本主打 东亚女性美学，整体风格拥有浓厚的 Instagram 与 小红书（Xiaohongshu） 氛围，适合喜欢清新、时尚、高质感人像的用户。📸🌸
🔥 更新内容
🤝 特别感谢芋头哥 @salttaro ！本模型融合了 Better Pussy & Poses V4，对整体效果帮助非常大。
🎨 加入了 2 个全新训练的 Moody Aesthetic LoRA，并从 ZImage 移植整合，让画面氛围更加出色。
🔞 增强 NSFW Prompt 遵循能力，提高生成稳定性与提示词响应效果。
NVFP4 格式
🚀 普通版 Krea2 V2 也将在近期发布，敬请期待！
感谢大家一直以来的支持！❤️
2026-06-25
YES - IT SUPPORTS ALL RACES - I AM JUST LAZY BUILDING A MULTICULTURAL PROMPT DB.
RH体验地址：https://www.runninghub.ai/zh-cn/post/2070402594421436417/?inviteCode=cptpovgo
🚀 Moody Krea 2 Mix V1 is HERE! 🎉🔥
After lots of testing, it's finally time to release the very first version of Moody Krea 2 Mix! 💫
This is the initial release—I'm testing the waters with Krea 2, but I have to say... this base model is absolutely amazing. 🤯✨
I've already got several improvements and future training plans in the works, so this is just the beginning! Expect even better quality and refinements in upcoming versions. 💪🎨
🔓 NSFW is fully unlocked.
Have fun, experiment, and let me know what you create! Your feedback will help shape future updates. ❤️
Enjoy! 🚀🔥
🚀 Moody Krea 2 Mix V1 正式发布！ 🎉🔥
经过一段时间的测试，终于迎来了 Moody Krea 2 Mix 的首个版本！✨
这是一次初始版本，先放出来试试水。不过不得不说，Krea 2 的底模真的非常惊艳，潜力十足！🤯
后续我已经规划好了几轮新的训练和优化，所以这只是个开始，未来还会继续提升画质、细节和整体表现，敬请期待！💪🎨
🔓 NSFW 已完全解锁。
欢迎大家尽情体验、多多出图，也欢迎分享你的作品和反馈，你们的建议都会成为后续更新的重要参考！❤️
祝大家玩得开心！🚀🔥
PS: Workflow in this image - 图里有工作流 I will be posting an official workflow shortly. 我晚点发配套工作流。
Description
FAQ
Comments (39)
fp32 -> 8gb ? & fp8 = 12 ?
fp32
Base Model
SD 1.5 ?
Relax guys, i had like 10 hours of sleep in total over the past 3 days
getting a substantial amount of black screens on the 4.0 nvfp4 version that all generate fine on 3.0 nvfp4.
I recommend you reboot your machine. i have had that after few nights of lora training
V4.0萌妹再出來，Moody真該好好睡著覺啦。我們都還來不及開箱，更新太神速了！
WOW, The hair finally become normal and no more messy hair
大佬太强了，更新的好快
更新速度真离谱。。。
hi there, you have a fatal error in your v4.0 where thread Array says Must NOT contain infs or NaNs. This is consistent even with a different workflow.
I'm not sure why it works with your current prompt, but the moment you change it, this occurs. Also I tried your v3.0 to double check, v3.0 works fine. So something is wrong with your v4.0 safetensor. 👍
If you use int8 you probably need to update comfyui to get it working
Yes version 4.0 (not int8) is broken and doesn't work
@Tribe28 Strange. It works fine for me.
You sure you dont have a outdated Comfyui?
Or maybe your download was corrupted. I have seen that multiple times before, that people just needed to download the checkpoint again.
@Tribe28 try downloading from my hugging face perhaps? Civitai has been giving people corrupted files lately
@catlover1937 I don't use Comfy but Wan2GP. Version 3 works flawlessly, when I try version 4 it just flashes and doesn't even start generating (it writes an error in the console). I tried downloading the version from huggingface, but that didn't help. Anyway, I tried sending this error to AI (Fable) and it provided me with some solution, I applied it to Wan2GP and now it works. 😀
Thank you very much for your update and the new version, it's great. 🥳
Anyway, the error in the new version 4.0 was related to this (according to the answer from AI), maybe it could be solved: The mmgp library (file offload.py) checks when loading the model that all checkpoint weights have the same data type (line assert model_dtype == dtype). Your new checkpoint has mixed data types - this is common for checkpoints prepared for ComfyUI (e.g. weights in fp8 and bias/scale tensors in bf16, or a mix of fp16 and bf16 after merging models). ComfyUI does not do any such check, which is why the checkpoint works there. In Wan2GP, this too strict assert fails.
@Tribe28 oh i know exactly what it is. A couple days ago, i sent a comparison post on how my FP8 was broken. So from V4.0 I started mixing with Official Krea2 FP8 instead of converting from BF16. so you (Wan2GP) prefers the broken version instead, lol
@catlover1937 Interesting. Well, I only have an Rtx 5060 Ti (16GB) and I'm using the nvfp4 version because it's really fast, I can accept a little bit of poor quality because I'll then run it through an upscaler (if the resulting image is worth it). But seeing your comparison, maybe I could try int8 convrot or Bf16 (my graphics probably wouldn't handle that) in Comfy, because I can't get int8 convrot (only FP8/nvfp4) to work in WanGP.
@Tribe28 You can run much larger models with offloading to system ram, with minimal speedloss. I use the int8 on my 12 GB 4070 ti with 10 seconds gen time for 1024x1024. On 50 series card i belive fp8 is faster then int8, but i would try out the larger size and do some simple timings to check if fp4 is any gain at all.
At least in comfy there is automatic offloading, and i would be surpriced if its not the same in Wan2GP.
@Loraholic 10 seconds? That's really fast, I can generate a 1024x1024 image in 22 seconds on FP4 on a 5060. On FP8 it takes almost twice as long. Yes, Wan2GP uses some advanced memory management, it will probably be similar to Comfy. But the 5060 is no miracle, it's the lowest card in the series, after all, it's quite weak even compared to the 4070. Anyway, I'll try a few more tests, I like the beautifully clean quality of BF16 that catlover1937 provided in his comparison. Thanks for your advice. By the way, you're doing great Lora. 😎
@Tribe28 nvfp4 does a great job already, I barely notice any difference if not put side by side
猫哥，希望在下一个版本加上年龄lora，这样我就不用找lora了（laugh）
先在此谢过猫哥
@sanyixianer 为啥要加年龄lora 你是觉得太老还是太嫩
这个更新速度，我只能说，大王勤政啊
the quality upgrade in v4 is outstanding. it gives sharp image really impressive.
AMAZING model. Easily solving acrobatic and gymnastic poses with minimum prompt tweaking! Also has great load of art styles in it.
The only downside - it tend to draw pussies on and crotch it seeing xD
V4 is a beast. Its creating europeans and dark skinned people very well as well, witch often is s problem with some checkpoints. My new goto model for sure.
And for 3000 and 4000 series GPUs, use int8. It takes 10 seconds to generate a 1024 image on my 4070ti, while fp8 takes 15 seconds.
can you share your workflow? also love your loras thanks for the hard work.
@necroryona My workflow is kinda small and simple. At least compared to some insane workflows out there. Some people have a twitsted view on what a simple workflow is.
But here is my "simple to understand" workflow, that i use for everything.
https://civitai.red/models/2692910/loraholics-simple-to-understand-workflow
Its for Z-image, but you can just copy the loaders from your exiting Krea workflow into my workflow and your are good to go.
@Loraholic thank you so much, i'll give it a try
@necroryona You can also try Moodys simple workflow which is optimized for this model. I recommend. Especially if you go tweakin couple things.
the first actually good image model
I don't know how you did it, but v4 is somehow even better, amazing work cheers!
Only one issue, I can't make the camera view from a low angle, any idea's?
There’s have to be some prompting technique involved, try refine your prompt using ChatGPT or grok, ask it to refine your prompt into a low angle shot, it helped me most of the times
You fixed actually the overcooking and higher base resolution sampling with this v4. Great job! 6/5 Got my platinum blonde now!
Btw the grid is there still if go to deep end and hair sometimes provides mourelle effect from sampler but should be fixable with glshader mask or double decoding through flux f.ex as aa technique. Im testing that at some point. I actually went updating all drivers and pushing new venv. Trying to see if something old is being wonky ass shite.
45s to 4k. Your simple workflow. 1.5 mpx starting resolution, skipping sd upscale, installed native (These are faster) seedvr2 nodes with int8convrot models = Great quality.
https://www.reddit.com/r/StableDiffusion/comments/1ustokp/comfy_native_seedvr2_workflow/
I was using json style prompts (some from Ideogram V4) for some content, and for complicated poses it seems that V3 is doing better than V4. Also some more obscure things like fisting were doing better with V3, than V4. What gives?
To be fair, i think this is normal there are always trade-offs of the goods and bads. You can stay with V3 if you are interested in some particular concept that V3 does better, there's nothing wrong with that!
NVFP4 is God!!
If I use anything else but your default prompt I just get noise, am I missing something?