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。
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 (67)
bf16 version please
You guys are beasts, you all have 26GB+ VRAM?
@catlover1937 It's not just the amount of VRAM, by not offering BF16 you're limiting the scope of usability across platforms.
@mrmrswiggly612 by far it seems like none of the creators on civitai released a BF16 version of their finetunes yet. I will look into this if a lot of people are requesting.
Long story short TLDR: I ran into OOM when i tried to compose the BF16 version
@catlover1937 Apple silicon doesn't support fp8
fp8 is sufficient for me
@catlover1937 I actually want to convert your models into int8 , and that requires bf16 source, if you can share it yourself it would be better of course. (fp16 is also ok I think if that makes any difference for you)
@xpnrt i don't have a BF16 version myself unfortunately, I ran into OOM when i tried to compose the BF16 version. I might need a more beefy setup
@catlover1937 so you do these merges on fp8 then ?
@xpnrt Yes, I did
yes fp8 doesn't work on any mac somehow. its structual. There is a patch but it works poorly , so basically fp8 excludes most mac users as FYI. thanks :)
@sanihands11225 I am actually so shocked that how many people are using mac for generation. BF16 is a 26GB model, even you have the most expensive mac studio it's not going to be efficient i think
@catlover1937 not bf16 specifically but nvfp4, and now int8-convrot is a fast and viable alternative , comfy themselves are releasing them in both if you check their official hf pages etc, both faster than fp8 for most people. Also regarding OOM's if you have the space just make the swap something like 200 GB. I have 16 gb rx 6800, 32 gb system ram and with 200 gb swap can even edit - merge full 40+ gb ltxvideo model with 3-4 loras at a time ...
@catlover1937 It's not a question of efficiency, it's a question of ability, can we use it or not? FP8, and all the rest including convrot are of no use to macOS, we need BF16.
I am geting plastic skin, a bit toward illustration and some deformation, with realism loras its a bit better, what might be the cause? how many steps, what sampler?
Please try out the paired workflow, i am also new to Krea2, but i guess it could be the combination of prompting and workflow
krea2 leans towards plastic skin in general, most people say
Do you have any comparisons between Krea2 and ZIT? I'd really like to hear an expert's assessment.
I am loving this. Nice job. I was so happy to see you put a model out there. Cant wait to see where you take this.
大佬方便上传Q8_0的gguf模型吗？
is it turbo model ? 8 steps is good enough?
Yes, please use workflow provided
i just modified your workflow, tried to use ZIT as face detailer model, so that i can use ZIT character lora while using Krea 2 for other part of the image. it work.
Yes!
Could you share the workflow? Or post a screenshot of it? Thanks.
@wyxzddsjj919 Seconded! Would love to see it.
I have my newest simple workflow available , technically, you just need to load the zimage model and connect to the any of the 2 detailers in the krea2 workflow and it should work
@catlover1937 3Q~MAX~
因为好奇才问的。 为什么只上传fp8呢？ 是因为bf16太重了吗？ 仅从生成质量来看，bf16好像更好，所以问一下，而且一直感谢好的作品。
Bf16 26gb 如果没5090用起来都是不实际的 哪怕毕竟4090 才24gb 。 但又说回来，目前来说bf16 我电脑没办法合成模型 ，会OOM错误。 如果真的很多人想要我考虑在云端合吧
因为我用的Mac系统，不支持原生FP8，所以用这个模型会报错，但是我尝试了您zimage系列的Pro，real，wild的fp8版本，我这里就可以运行了，所以想请教您一下，您zimage系列的fp8，和krea2的fp8有什么区别吗？
mac系统正常应该不支持FP8。我也是用mac，我都是用FP16.否则会报错
@gagayoge2318 就是官方的FP8融合的，zimage的FP8是自己从BF16做的
@gagayoge2318FP16也可以，Q8_0的GGUF也可以，我测试过好像只有标准FP8的float8_e4m3fn格式会通过CPU反量化到FP16，再回到GPU运行，FP8_scaled不能运行，具体情况我还在查。
@catlover1937 官方的FP8是带有动态 Scale 因子的吗？
求gguf版
同求
@shangcharels1616633 好像有个大佬做了 在c站转发了
to controll materials,light and achieving accurate colors has become more difficult in the Krea 2 model.
大佬可以练个FP4的嘛~！
i think to achieve these variants, i need to first get BF16 to work, but my rig is bit too shitty to handle it. I am looking into cloud solutions.
INT8 version pls!
i think to achieve these variants, i need to first get BF16 to work, but my rig is bit too shitty to handle it. I am looking into cloud solutions.
@skyrimer3d Do you have any working code that turns BF16 to INT8 for Krea? if you do can you please share?
int8 version would be really nice please add it (the speed difference is huge for 4xxx series gpu) compared to fp8
i think to achieve these variants, i need to first get BF16 to work, but my rig is bit too shitty to handle it. I am looking into cloud solutions.
Do you have any working code that turns BF16 to INT8 for Krea? if you do can you please share? @AI_Aurora
@catlover1937 i dont know much about converting but i found this. he used converted using silveroxides/convert_to_quant (ctq CLI) with ComfyUI-compatible metadata.
Conversion Details
All variants converted with:
--simple (round-to-nearest, no learned rounding)
--comfy_quant (ComfyUI-compatible layer names and metadata)
--save-quant-metadata (quantization info embedded in file)
--low-memory (reduced peak RAM usage during conversion)
--flux2 profile
ConvRot variants additionally use Hadamard rotation (--convrot) prior to quantization to reduce outliers and improve accuracy.
https://huggingface.co/Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8
@catlover1937 this model is quite vram dependent and most of the basic user dont have more tha 8gb vram. i have 4060 8gb and the fp8 model takes 35 to 45 sec and int8 only takes around 20 to 25 sec. huge difference. but that is only available for base krea2t. would be really helpful for everyone if you could do your model int8 version since your model gives one of the best results.
@AI_Aurora I tried this earlier today already. the int8 version is not recognized by comfyUI
@catlover1937 Conversion Commands
Standard INT8 (Row-wise)
bash
convert_to_quant -i krea2_turbo_bf16.safetensors \ --int8 \ --block_size 128 \ --comfy_quant \ --flux2 \ --scaling_mode tensor \ --simple \ --save-quant-metadata \ --low-memory
ConvRot INT8 (Better Quality)
For the best quality INT8 variant (recommended for RTX 30xx), add --convrot:
bash
convert_to_quant -i krea2_turbo_bf16.safetensors \ --int8 \ --block_size 128 \ --comfy_quant \ --flux2 \ --scaling_mode tensor \ --convrot \ --simple \ --save-quant-metadata \ --low-memory
Flag Explanations
Table
FlagPurpose
--int8 (Target format: INT8)
--block_size 128 (Block size for quantization)
--comfy_quant (ComfyUI-compatible layer names and metadata)
--flux2 (Use the FLUX.2/Krea2 quantization profile)
--scaling_mode tensor (Per-tensor scaling (use row for per-row))
--simple (Round-to-nearest (no learned rounding optimization))
--save-quant-metadata (Embed quantization info in the output file)
--low-memory (Reduce peak RAM usage during conversion)
--convrot (Apply Hadamard rotation before quantization to reduce outliers)
@catlover1937 do you have latest, latest, latest confyUI? Native INT implementation is fresh, maybe one day old.
@catlover1937 int8 take separate custom node like gguf named ComfyUI-INT8-Fast
@AI_Aurora you dont need custom nodes anymore, new update give us NATIVE INT8 support.
@derispan6661071 yeah this is what I read as well. But it just won’t load
@catlover1937 look at latest Winnougan youtube video, he also has his hugginsface page.
@derispan6661071 Yep, i already tried CTQ, no luck
@catlover1937 sorry mate, I don't know how I can help you.
@derispan6661071 perhaps it actually works with custom int8 nodes, i only tried with the default comfyui loader as the newest release advertise that they support INT8, but maybe its on them
@catlover1937 yes try with this https://github.com/BobJohnson24/ComfyUI-INT8-Fast i get 2.5s/it in int8 version using this node but 7s/it plus with the default comfyui native node and normal fp8 3.5 to 4s/it so huge speed boost for int8 version with this custom node
Thank- you, outstanding work. I have trained a Lora with krea2_raw and it works with basic Krea2 turbo no issues. I can't get it to work with this checkpoint and your basic workflow. I really have to increase strength just to see her and then it comes out terrible. Any tips or am I training her wrong? Thankyou
Hey, i mean the problem could be on ME, if it works fine on base but not mine. I am also new to Krea2, but more refinements will be added to enhance lora compatibility
It was my fault. I retrained again on Raw with more steps and applied it to your V2, all I can say is Thank you. Amazing model, sticks to the prompts, complete uncensored. I loved your zturbo moody real mix but this is my new go to model now. appreciate the hard work you do.
To release more variants of the model, i need to first get BF16 to work, but my rig is bit too shitty to handle it. I am looking into cloud solutions.
哪个配置玩不动战地16噢
We'd like a BF16 version please. The other ones are no good for us.
@mrmrswiggly612 i hear ya, but i need my codex reset and give it another try