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！
Ko-fi: ko-fi.com/catlover1937
Please do not redistribute without consent. 如非作者同意禁止搬运。
2026-05-14
V4 distilled model is available now. V4 蒸馏模型放出。
2026-05-12
🚀 Hello everyone — the model you’ve all been waiting for is FINALLY here - MOODY WILD V4.0! 🔥
This is a MASSIVE leap from Wild V3, and unbelievably… the generation quality now even surpasses ZImage Turbo models 🤯
Honestly, I’m still shocked myself by how good the results turned out — and I couldn’t be happier with it.
Sorry for the slow updates on my other models lately 🙏
Most of my time has been spent training, testing, and merging this project to perfection.
And trust me… this is NOT like any other ZIB undistilled model.
Give it a try — I’m confident you’re going to LOVE it 💥
Huge thanks to @salttaro ❤️
for helping with testing and generously letting me use his machine so I could get this release out to you all today!
📢 : I’m currently wrapping up the distilled model, and it’ll be with you very shortly ⚡
The promised excellence is coming — and trust me, it’s absolutely worth the wait 🔥
✨ What’s New:
Insanely stunning results (comparable to ZIT! 🤯)
Outstanding lighting & spatial understanding 💡
Massive enhancement to NSFW capability 🔥
Greatly reduced bad body anatomy 🧍
Can now be used standalone with my ZIB Simple Workflow ⚡
Also good to use with my ZIB + ZIT workflow.
You know the drill 😎
Stay safe & have fun generating! 🎨✨
🚀 大家久等了——你们一直期待的模型终于来了 MOODY WILD V4.0! 🔥
这是一次相较于 Wild V3 的巨大飞跃，甚至不可思议的是……生成效果已经超越了 ZImage Turbo 模型 🤯
说实话，就连我自己都被这次的效果震惊到了，真的非常满意！
最近其他模型更新比较慢，真的抱歉 🙏
因为这段时间我几乎把所有精力都投入到了这个模型的训练、测试和融合上。
而且相信我……这绝对不是普通的 ZIB Undistilled 模型。
亲自试试看吧，我相信你一定会爱上它 💥
特别感谢 @salttaro ❤️
帮忙测试，并且借给我机器使用，才能让我尽早能把它发布给大家！
📢 :
我目前正在收尾 Distilled 版本，很快就会带着它和大家见面 ⚡
承诺过的极致效果马上到来——相信我，它绝对值得等待 🔥
✨ 更新内容：
效果惊艳到离谱（已经能媲美 ZIT！🤯）
极强的光影与空间理解能力 💡
NSFW 能力大幅增强 🔥
大幅减少错误人体结构问题 🧍
现在可直接搭配我的 ZIB Simple Workflow 单独使用 ⚡
同时也推荐使用我的 ZIB + ZIT workflow.
老规矩 😎
注意身体，玩得开心，尽情生成吧！🎨✨
2026-04-19
EDIT 2: 🤷♂️
Don’t ask why there are so many downloadable versions—they’re most likely the same thing. Just CivitAI being CivitAI. 别问为什么有这么多可下载的模型或者为什么无法下载——它们很可能是同一个文件，CivitAI日常抽风神仙也救不了。
EDIT 1: 🛠️
TIPS: You can use the distilled model with my simple workflow. 小贴士：蒸馏模型可以搭配我的简易工作流一起使用。而且还有EASON留下的彩蛋哦 🥚😮
V3 Distilled Model Released! 🚀
The V3 Distilled model is now live! This version is a fine-tuned iteration based on the V3 x EASON base model, where I explored several experimental ideas to push visual quality further. Thanks again to @EasonnLi !
🔥 What’s new:
Ultra-detailed skin textures – significantly enhanced micro-details for a more lifelike appearance
Improved realism – better lighting response and natural rendering
Expanded pose capability – performs well across a wide range of poses (both SFW & NSFW)
More expressive outputs – stronger visual impact overall
⚠️ Known trade-offs:
Higher error rate – due to aggressive tuning, generation stability may be slightly reduced
Text distortion issues – inherited (likely) from the EASON base, especially noticeable in text-heavy images
This release is more experimental compared to previous versions, aimed at pushing boundaries rather than pure stability.
💬 I’d love to hear your feedback—your support and testing help shape future updates! 💴
V3 蒸馏模型正式发布！🚀
V3 Distilled 模型现已上线！本次版本基于 V3 x EASON Base 模型 进行微调，并加入了一些实验性的优化思路，进一步提升画面表现力。再次感谢@EasonnLi的辛勤付出。
🔥 更新亮点：
极致皮肤细节 —— 皮肤纹理更加细腻真实
更强写实感 —— 光影表现更加自然
姿态表现增强 —— 支持更丰富的姿势（SFW & NSFW 均有良好表现）
整体表现力提升 —— 画面更加有冲击力
⚠️ 已知问题：
错误率有所提升 —— 由于调校较激进，生成稳定性略有下降
文字畸变问题 —— 可能继承自 EASON 模型，在含文字图像中较明显
这个版本更偏向实验性质，目标是探索上限，而非单纯追求稳定性。
💬 欢迎反馈你的使用体验，你的支持对后续优化非常重要！💴
2026-04-17
Edit 1: 你们着急的话 可以先用fun 的8步lora 用着, 我的蒸馏版应该要明后天才有空发。use 8 steps Lora from the fun team to speed up generation of wild v3 base if needed, my own distillation is still work in progress
https://civarchive.com/models/2362961/z-image-fun-distill-lora
Hiya 👋 Moody Wild V3 BASE is here!
This time, it’s a collaboration with EASON’s NSFW ZimageBase V3 model — definitely go check it out. It delivers insanely detailed NSFW results and is seriously underrated.
For V3, I’ve merged Moody Wild V2 with EASON’s base, then layered in several of my Moody-style LoRAs from the PORN and REAL series to push both realism and intensity further.
I hope you’ll enjoy this one 🔥
✨ What’s New
🌍 Less bias, better diversity — strong character results across all races
⚡ More standalone usability — performs great even without speed-up LoRAs
🔍 Enhanced fine details — especially in genitals and anatomy
💡 Improved lighting & composition — more natural photo structure
🎨 Moody style preserved — same signature look, now more refined
🧩 Usage
You can use this model:
As a standalone (40 STEPS CFG 4.0)
With your favorite speed-up LoRA
Or inside my ZIB + ZIT workflow
(Moody distilled version coming soon after further fine-tuning 👀)
Enjoy and have fun 💫
Your support is always appreciated ❤️
Hiya 👋 Moody Wild V3 来了！
这次是与 EASON 的 NSFW ZimageBase V3 模型 的合作 —— 强烈建议大家去看看，这个模型在 NSFW 细节表现上非常夸张，但却一直被低估。
在 V3 中，我将 Moody Wild V2 与 EASON 的 base 进行了融合，并叠加了多个我在 PORN 和 REAL 系列中的 Moody 风格 LoRA，进一步强化真实感和表现力。
希望你们会喜欢这次的版本 🔥
✨ 更新内容
🌍 更低偏差，更强多样性 —— 各种族角色表现更稳定、更优秀
⚡ 更强单模可用性 —— 即使不搭配加速 LoRA 也能很好使用
🔍 细节大幅增强 —— 尤其在人体结构方面
💡 光影与构图优化 —— 更自然的照片结构表现
🎨 保持 Moody 风格 —— 熟悉的味道，更进一步打磨
🧩 使用方式
你可以这样使用这个模型：
作为独立模型使用 (40 STEPS CFG 4.0)
搭配你喜欢的 加速 LoRA
或结合我的 ZIB + ZIT workflow 工作流 使用
(Moody 蒸馏版本正在进一步微调中，很快发布 👀)
祝你玩得开心 💫
你的支持一直都非常重要 ❤️
WARNING: USING UNDISTILLED BASE MODEL (V1 AND V2 ONLY) ALONE IS NOT RECOMMENDED, UNLESS YOU KNOW WHAT YOU ARE DOING, INSTEAD USE ZIB + ZIT workflow
注意：不建议直接使用V2以及之前未蒸馏50步模型，除非你知道你自己需要什么, 未蒸馏模型需配合ZIB + ZIT workflow
2026-03-19
GGUF versions provided by @CH11442 , please show your thanks to him directly!
New ZIB + ZIT workflow V4 available！
2026-03-18
Moody Wild V2 – Release Notes
Alright, I gotta start by saying this: Zimage is ridiculously hard to train.
Maybe it’s my slow hardware, maybe it’s just the nature of Zimage itself, or maybe I’m just slowly losing my patience 😅.
For the past week I’ve been trying to push out a fine-tune that’s actually better than V1, but I kept running into weird overtraining issues when trying to improve character quality. Random things like:
glowing hair bleeding onto female chests
random ugly goosebumps on skin
greasy-looking hair
weirdly distorted nipples
and other cursed artifacts
At some point I honestly thought about just giving up on this version. But Moody Wild has always been my experimental playground, where I try out the newest ideas and training tweaks. So I decided to keep pushing and see what happens.
So please bear with me if this release isn’t perfect. That’s kind of the spirit of this model.
Hopefully V2 brings:
✨ Better looking characters
✨ Fewer malformed poses
✨ Slightly more stable generations overall
Sorry for the rambling — but I hope this version ends up being a step forward from V1.
As always, feedback helps a lot. Enjoy experimenting with it!
先说一句实话：Zimage 真的太难训练了。
可能是因为我的硬件太慢，也可能是 Zimage 本身的特性，或者也可能只是我自己的耐心正在慢慢被磨没了 😅 但无论如何，试错成本实在是太高了，重练一次就至少一天一夜
过去这一周我一直在尝试做出一个比 V1 更好的微调版本，但训练过程中一直出现各种奇怪的过拟合问题。本来是想把角色画得更好看，结果却不断出现一些奇怪的情况，比如：
头发发光，甚至“溢出”到女性胸口
皮肤上出现随机又很丑的鸡皮疙瘩
头发看起来油腻腻的
乳头形状诡异甚至畸形
以及其他各种奇怪的训练伪影
老实说，有一段时间我真的想直接放弃这个版本了。
不过 Moody Wild 一直都是我的实验模型，很多新的训练思路和尝试我都会先在这个模型上测试。所以最后还是决定继续折腾下去看看能不能有所改进。
所以如果这个版本没有达到大家的预期，还请多包涵，这本来就是一个偏实验性质的模型。
希望 V2 相比 V1 能带来：
✨ 更好看的角色表现
✨ 更少的畸形姿势
✨ 整体生成稳定性稍微提升一些
抱歉写得有点碎碎念，不过还是希望这个版本能比 V1 有所进步。
如果有任何反馈也非常欢迎告诉我，祝大家玩得开心。
有条件的小伙伴尽量选BF16版本下载
Please try to use BF16 version if your hardware allows
2026-02-22
GGUF Verions Provided by @CH11442 , please show your thanks directly @CH11442 !
WARNING: USING UNDISTILLED ALONE IS NOT RECOMMENDED, UNLESS YOU KNOW WHAT YOU ARE DOING, INSTEAD USE ZIB + ZIT workflow
注意：不建议直接使用未蒸馏50步模型，除非你知道你自己需要什么, 未蒸馏模型需配合ZIB + ZIT workflow
In V1.0, there are two versions available: Distilled and Undistilled.
The Distilled version runs similarly to turbo models, performing well at CFG 1 with around 10 steps.
The Undistilled version, on the other hand, requires a CFG of around 4 and approximately 50 steps to fully render.
Please note that the Distilled and Base (Undistilled) models will produce noticeably different artistic styles in their final results.
The Undistilled version is not intended to be used on its own, as it is not yet artistically refined (at least at this stage). This is also why most creators choose to release only the distilled version.
So why include the base version?
Training purposes – It can be used for training. However, I still recommend training on the vanilla base model, since the moody version will continue evolving, and it’s not practical to retrain for every new iteration.
Workflow completion – Most importantly, it completes the “Moody Ecosystem” within the ZIB + ZIT workflow, helping deliver the ultimate realistic xxx graphics.
I will provide more detailed instructions explaining why ZIB + ZIT are both needed in the workflow. Download links will also be included, so you can freely mix and match your preferred models.
在 V1.0 中，一共提供两个版本：Distilled（蒸馏版） 和 Undistilled（未蒸馏版 / Base 版）。
Distilled 版本 的运行方式类似 Turbo 模型，在 CFG 1，10步 的情况下即可获得良好效果。
而 Undistilled 版本 则需要 CFG 4 ，并且大约 50 步 才能完整出图。
需要注意的是，Distilled 与 Base（Undistilled）版本在最终生成的艺术风格上会有明显差异。
Undistilled 版本并不建议单独使用，因为它目前在艺术性上还没有完全打磨成熟（至少在现阶段如此）。这也是为什么大多数创作者通常只发布蒸馏版本。
那么，为什么还需要 Base 版本？
用于训练 —— 可以作为训练用途。不过仍然建议在原版 base 上进行训练，因为 moody 版本会持续迭代更新，你不太可能为每一个版本都重新训练。
完善工作流生态 —— 更重要的是，它可以补全 ZIB + ZIT 工作流中的「Moody 生态」，以实现最终极的写实 xxx 画面表现。
后续我会补充更详细的说明，解释为什么在工作流中需要同时使用 ZIB + ZIT。同时也会在工作流中附上下载链接，方便大家自由搭配自己喜欢的模型组合。
2026-02-17
V0.2
CFG 1.0 | 10 STEPS.
dpmpp_2m_sde | beta
Refinements over the previous version to improve stability. The next release will probably be a full release of a 5 Steps distilled model if I can get it to work out. Appology for the slow update as my rig is getting slow and dated.
在上一版本的基础上进行了优化和调整，以提升稳定性。
如果一切顺利，下一次发布应该会是一个包含一个 5 步蒸馏模型的完整版本。
很抱歉更新比较慢，因为我的设备已经有些老旧吃力了，运行速度也比较慢了。
2026-02-11
V0.1
CFG 1.0 | 8 - 12 STEPS.
dpmpp_2m_sde | beta
A very immature experimental Zimage Distilled model focuses on drawing ultra realistic uncensored scenes.
PS. I may or may not continue developing this model based on user feedbacks.
這是一個非常不成熟的實驗性 Zimage Distilled 模型，專注於繪製超逼真的解锁場景。
附註：我可能會也可能不會繼續開發這個模型，這取決於使用者的回饋。
Description
FAQ
Comments (46)
for the double ksampler, how should i split the steps between the first and 2nd ksampler on your workflow for the 40 step version? Because no matter what I do i get a black image
Please try my Zimage base workflow
黑屏或者有色块，可以试着关闭乃至卸载 Sage Attention 来解决问题。
@catlover1937 I did, but it did not work
the distribution of samplers is not critical, it works in a wide range, i go for about 3:4, if the result is completely black something else goes wrong.
@fierceys turn off sage attention there’s a bug with comfy and ZiB models
@catlover1937 turning it off worked, but it's very annoying. Thank you
@fierceys have two start bat scripts, it’s simpler than you thought
v4延续了v3的文字差和不能风格化😂这两点还得是看v2
Thank you for sharing!
Sadly the output from v4 distilled fp16 are very blurry, like is missing the distill Lora. i'm using the same config as your examples: 9 steps, 1 cfg, dpm++ 2m sde, beta.
have u tried my workflow?
V4 seems to be very heavy on Asian only faces.
yup, v3 and v4. I just switch back to v2 right now for non-Asian
you can just use the desired ethnicity in your prompt!!!
支持zimage base训练的人脸lora吗？为什么加上lora之后没有效果并且很模糊
是支持的，排查一下是否是lora问题还是模型兼容度问题。可以用我的人物lora 测试https://civitai.red/models/2629822/moody-gal-dori
@catlover1937 佬，按照工作流中使用FaceDetailer的方式感觉还可以。不过我有个新的问题,使用ZIB/ZID 4.0 和提供的两步采样工作流。每一步的模型 步数和cfg怎么设置
@nongli9100113 我zib zit工作流的默认就是最佳设定 加载模型就可以直接使用了
大哥，救命啊，你这个模型哪哪都好，就是nipples严重扭曲变形啊！！能不能再改进一下，后期的重绘太麻烦了！
真的吗，谁的工作流
@catlover1937 sorry,没说清楚，是在近景人物特写的情况下会出现这种情况，中远景不会有问题。在SD放大部分会严重放大这一缺陷，我在这个节点更换模型解决的这个问题，唯一就是需要在一个工作流中加载两个模型
我经过几轮测试，确定了只有在画面中同时出现Large breasts和Pussy时会出现nipples扭曲变形的状况，按图片的情况来看应该是某个集成的Lora把nipples和clitoris混为一样物体了，因为nipples扭曲的皮肤形态都是类似clitoris的形态。请在下一轮改进中调整一下。
看来还是人体结构打标的问题，是用蒸馏模型发现的吗，有可能是蒸馏模型冲突导致的
Hello very great work buddy I played with your base v4 model for a while. Found out while using character lora from z-image or even training character lora in ostris ai-toolkit with same settings as z-image lora on your model provides quite terrible results compared to training on original z-image. It it about prompting your model or something else? I use one ksampler workflow and I generate in like 1080p quality without upscaling. As I said z-image is fine while your model disappointed me here.
Never ever train on custom models unless you know what you are doing. Imagine if i release V5 tomorrow, you V4 loras will be completely useless. Training on custom models requires much more experience and with many trials and erorrs, so I can't advise you much on it.
In terms of using character loras, do you use my character loras? what character loras are you using, and what's the strength you put to them? Let's troubleshoot together
@catlover1937 I use my own lora buddy. I don't mind skipping a version v5 and train again on v6 I just need good NSFW model for my ai influencer. I know training a lora on custom modes is not so easy as on official model, but it's mostly "garbage in, gargage out" and my dataset is good. Okay so I've tried a few different approaches and I figured out it requires detailed prompt while loading lora the results are okay then, so about 900+ characters while loading lora. I used 1.0 and 0.9, never really go below that. What values do you recommend? Of course we talk about realism, just wanted to clarify.
@wezaluketek345 it depends hugely on your Lora rank and steps. It’s different for each Lora, for my loras I can push it to 1.5 without much trouble
@catlover1937
linear: 32 linear_alpha: 32 conv: 16 conv_alpha: 16
Ended at ~ 5000 steps, slightly too much, cuz the face and body was fine way sooner, but lower checkpoints getting same results but it may be just your model style I dislike, I don't really know bro. It's surely one of the best open NSFW model so great work
@wezaluketek345 @wezaluketek345 @wezaluketek345 yeah, i can't tell for sure, it could be the dataset too. my loras were trained on base and it works fine, so maybe you can give my loras a try and see if you like them. Or i can make the lora for you for a small fee.
AMAZING MODEL !!!!
People need to learn and understand how to use specific models for specific tasks. I see that many people who use checkpoints don't know how they work, how to use them, or that they aren't general-purpose models. Also use the correct workflows
Dear experts, does the Moody Wild Mix model use UNET loader or checkpint loader?
unet
I spent 2 days trying to use this model (V4.0 Distilled 10 steps), but unfortunately, most of the results I get are blurry (not sharp). I don't have this issue at all with other Z-image models, whether they are base or distilled versions, and I'm definitely not a beginner when it comes to ComfyUI. It's really a shame.
I am sad to hear about that. All my demo images were uploaded with workflow embeded, you can try them out if you want. Maybe you can try to reproduce them first, if you think the demo images were bad too, then maybe this model is just not for you.
@catlover1937 I get the impression it’s down to the wording of the prompt, which is much stricter – have you changed anything? I’ve tested some prompts from the user "rockiecxh" and the images are fine. I’ve tried asking the AI to generate similar prompts for me, but even it can’t produce prompts that are clear.
@catlover1937 Here are three prompts I’ve tried which give me blurry images – could you have a go with them?
Prompt 1:
The image is a photograph of a young woman with medium-length, dark brown hair and light olive skin. She's sitting on the carpet in front of a TV, wearing a black t-shirt and white sweatpants with white socks. She's making a heart shape with her hands near her face and smiling directly at the camera. The lighting is artificial and bright, highlighting her face while the background remains dark and blurred. The camera angle is slightly below eye level, creating a medium close-up shot that uses the rule of thirds for composition. Her age appears to be in her twenties. The depth of field keeps her in sharp focus while the background is out of focus.
Prompt 2:
Low-angle cinematic action shot of a female knight in the heat of a chaotic medieval battle, captured at the peak of a powerful horizontal sword swing. She wears intricate, battle-scarred plate armor with silver engravings and visible scuffs, reflecting the flickering orange glow of nearby embers. Her expression is one of intense focus, jaw set, as she rotates her torso to drive the blade. Dramatic motion blur follows the arcing trajectory of her heavy steel broadsword, which slices through a cloud of thick, volumetric dust and a spray of brilliant amber sparks. The background is a hazy, desaturated battlefield with out-of-focus silhouettes of clashing soldiers and tattered banners. Harsh, directional sunlight pierces through rising smoke, creating high-contrast highlights on the polished metal edges. Shot on 35mm lens, f/2.8, shallow depth of field, with raw textures of flying grit, sweat, and oxidized metal.
Prompt 3:
A young woman with wavy light-blonde hair, softly lit by warm golden light from the side. She wears thin round-framed glasses and a simple white t-shirt with a subtle red detail on the chest. she have cute smile on her face, One small stud earring is visible. Her gaze is directed upward and slightly away from the camera, expression calm and thoughtful. Background is softly blurred in cool tones, contrasting with the warm highlight on her hair and face. Lighting is cinematic â soft key light with gentle rim glow. Mood: dreamy, introspective, serene.
@Light_x02 Could it be the depth of field causing the problem? I will be keep this in mind when I work on V5.
@catlover1937 That’s possible. Did you have the same problem as me with the prompts I gave you?
@Light_x02 No, not really. Wasn't the highest res, but clarity seems ok (tesed with with my zib workflow).
@catlover1937 Have you tried using just one ksampler? Because 80 per cent of users only use one ^^
@Light_x02 Really? I thought people would use what it’s paired with. I know this is a bad comparison, but it’s like using Apple AirPods with Android or Windows devices. You’re probably right that 80% of people use Android and Windows. And I am sure testing the AirPods on non-Apple devices is very limited, if you ever tried to make them pair lol.
But yeah I get what you are saying, I will be mindful of this when making the next versions. Thanks for reporting. 🫡
@catlover1937 You're welcome :) I'll have a look at incorporating the two Ksamplers I'll be using into my workflow, and if I get some good images, I'll be happy to share them ;)
Where do I find thi s missing model?: ComfyUI\moody-wild-v4-undistilled-40steps-cfg4_00001__clean.safetensors
that's the V4.0 BASE 40STEPS CFG4
用了看下来，v4的10步加速模型，单采样器出图会模糊糊的，请问猫哥是用双采样吗？v4的单采样直出图效果不怎理想
是的呢，强烈建议用我的工作流，