SUPER Z image turbo rebuild 淡妆浓抹 V3.0 菁彩发布

20260210更新lora训练适配器

注意，是淡妆浓抹 3.0 模型的的lora训练专用，重点是diffusion，diffusion是新的训练才是新的，适配器是adapter，Rebuild版底模支持在此适配器上训练，可在aitoolkit中替换并微调自己专属lora。注意，不同于官方底模，是专用于本模型底模的训练适配器，感谢@Laushine制作，请改扩展名为ZIP后解压使用。

个人微调版 Z image turbo rebuild 淡妆浓抹 3.0发布

20260210补充FP8版

Z TURBO REBUILD , 更好的色彩光影表现，更细腻的纹理细节呈现。

比 Z image turbo 更多的艺术风格支持，比 Z image base 更快的生成速度。

CFG:1，10~15步，

euler+simple 采样器调度器

善用SeedVarianceEnhancer等节点，增加画面的多样性表达
优化了一些艺术风格的支持与二次元动漫的表现。

Nunchaku量化版 by tonera(感谢）：

https://modelscope.cn/models/tonera/Beyond_Reality_Zimage_v2_svdq/summary

NVFP4量化以及采样器搭配 by miccodadada（感谢）：

本人实测，AuraFlow3.5 - 5，首选euler a+FlowMatchEulerDiscreteScheduler,其次是ZImagePowerNodes自带的采样器，然后是DEIS_2M + BONG_TONGENT，追求泛用性的话是Sa_solver+FlowMatchEulerDiscrete,人像的话也可以试试ER_SDE+(sgm_uniform/FlowMatch)。

我自己制作了它的NVFP4混合量化，FEED FORWARD 123层采用NVFP4，其余注意力采用FP8 Per-Tensor量化（目前Comfyui暂时不支持BlockWise量化），质量下降在可接受范围内。在5070上，9步采样时间大概2.8-3.6秒。https://huggingface.co/lieding1994/zimage-serverless/blob/main/BEYOND-REALITY-BF16_learned_fp8mixed_tensor.safetensors

欢迎各位试用
最后，团队继续招新，我一手搭建起Z image数据生产团队，目前因数据需求不断增加，Z image数据生产团队开启招新，坐标杭州某大厂，需要全职到岗。有意者加WX：Aptshtiso

SUPER Z 淡妆浓抹 V2.0 菁彩发布

解决了一些已知问题

优化了纹理与细节表现，继续优化模型的高频细节表现。

可能拥有目前所有Z模型中，最佳的直出纹理细节表现能力，同时对胶片等摄影风格支持更佳。

同时尝试优化一些艺术风格的支持与二次元动漫的表现。

常规升级，欢迎试用，请多多返图~

感觉比淡妆浓抹1.0版进步还是较为明显的（放大看细节表现）

为了获得多样性，请尝试SeedVarianceEnhancer节点

CFG:1.2

采样步数：10~15步

最后，团队招新，我一手搭建起Z image数据生产团队，目前因数据需求不断增加，Z image数据生产团队开启招新，坐标杭州某大厂，需要全职到岗。有意者加WX：Aptshtiso

欢迎加入我们Z image数据生产团队！

https://www.modelscope.cn/models/Nurburgring/BEYOND_REALITY_Z_IMAGE

https://civarchive.com/models/1090420?modelVersionId=2581368

https://huggingface.co/Nurburgring/BEYOND_REALITY_Z_IMAGE

————————————————————————————————————————

SUPER Z Effortless Makeup - V2.0 - Vibrant Color Release

Resolved some known issues

The texture and detail presentation have been optimized, and the high-frequency detail presentation of the model is also being further improved.

It may possess the best direct rendering texture detail performance capability among all current Z models, and also offers better support for photographic styles such as film.

At the same time, efforts are made to enhance the support for various artistic styles and to improve the presentation of 2D anime.

Regular upgrade. Welcome to try it out. Please share your feedback!

The effect seems to have improved significantly compared to the previous version (viewing it in detail reveals this).

To achieve diversity, please try the SeedVarianceEnhancer node.

CFG: 1.2

Sampling steps: 10 to 15 steps

Finally, the team is recruiting new members. I have established the Z image data production team by myself. Currently, due to the increasing demand for data, the Z image data production team has launched a recruitment drive. The location is a major company in Hangzhou. The position requires full-time employment.

Those interested can add WeChat: Aptshtiso

Welcome to join our Z Image data production team!

————————————————————————————————————————

SUPER Z 浓墨重彩 V1.0 菁彩发布

祝大家2026年元旦快乐~

模型如其名，浓墨重彩

注意，浓墨重彩版适合网红风格人像图片的生成，如果不喜欢这个风格，请使用浓妆淡抹1.0版。浓妆淡抹1.0版仅对纹理细节与质感做了优化。

这是一个网红偏好方向的微调模型，目前的浓墨重彩版本相比浓妆淡抹1.0版，继续优化纹理细节表现，纹理清晰，细节丰富，光影效果更加浓烈，搭配风格提示词效果较好，色彩表达大胆，偏胶片又不相同，略带磨皮效果，适合时尚写真类图片或社交网络风格图片。
同时尝试提升艺术化风格的表达能力，只是测试下来比较失败，艺术风格的支持还是不好。后续或许会专门出艺术风格支持良好的模型。

所以这是一个偏重口的高精度人像模型，请悉知

为什么搞这个模型？因为我发现各个讨论群中大量发布网红风格图，有些图很好看，但是细节表现并不好。这个模型如果你使用得好，你可以出很精彩的网图风格图片，如果控制不当，你会很失望。非模型之罪，风格取向不同。如果你觉得不好，尽管骂我~

后续或许模型会分为两个分支：

浓妆淡抹，以提升纹理细节表现为主

浓墨重彩，以提升色彩光影表现为主

有用户反馈模型会有AI感，全看着自己的控制了哈哈哈

可以尝试在提示词前加一些风格描述，比如：

富士胶片风格

柯达胶片风格

徕卡镜头风格

蔡司镜头风格

电影风格

黑白默片风格

拍立得风格

ins风格

小红书风格

。。。。。。

你还可使尝试各种风格，或许会有不一样的惊喜

不要在意下载的模型名字还是浓妆淡抹，就是浓墨重彩，忘记改名了哈哈哈

SUPER Z - A Majestic and Colorful Grand V1.0 Release

Wishing everyone a happy New Year's Day in 2026!

The model is named as such, with rich and vivid details.

Note: Version Majestic and Colorful Grand is suitable for generating images of popular style portraits. If you don't like this style, please use Version heavy and light makeup 1.0. Version heavy and light makeup 1.0 only optimized the texture details and texture quality.

This is a micro-tuned model for the preferences of online influencers. Compared to version 1.0, the current version continues to optimize the representation of texture details. The textures are clear, the details are rich, the lighting effects are more intense, and the combination with style prompts works well. The color expression is bold, it leans towards film but is different from it, with a slight touch of smoothing effect. It is suitable for fashion photo-like images or social network-style images.

At the same time, an attempt was made to enhance the expression ability of the artistic style, but the test results were not satisfactory. The support for artistic styles is still not good. In the future, a model with good support for artistic styles may be specially developed.

So this is a high-precision portrait model with a pronounced facial feature. Please be informed.

If you handle it well, you can create very impressive web-style pictures. But if you don't control it properly, you will be very disappointed. It's not the fault of the model; it's due to different style preferences.

You can try adding some style descriptions before the prompt words, for example:

Fuji Film style

Kodak film style

Leica lens style

Zeiss lens style

Film style

Black-and-white silent film style

Positivie style

The Ins style

Little Red Book style

。。。。。。

You can also try various styles, and perhaps you will have some unexpected surprises.

SUPER Z 1.0或许是现在最佳的人像摄影 Z image 模型

注意最下面的参数设置。

好久不见，甚是想念

BEYOND REALITY系列模型断更已经半年多，主流开源绘图模型也已经快速迭代到了Z时代。

Z IMAGE给大家带来了足够的惊喜，超快的速度和超真实的画面，社区中对edit的呼声更是持续不断。

我知道你们很着急，但是先别急，先来看看社区微调版的 Z image 哈哈哈

BEYOND REALITY Z IMAGE 1.0 淡妆浓抹总相宜，这是一个基于人像数据集微调lora合并而来的模型，延续BEYOND REALITY系列模型的特点，高清晰、高美学。Z IMAGE目前的问题之一就是高频信息的表现不佳，体现在画面上就是纹理的细节不好。因此训练了一个lora来提升模型的纹理质量。本想发布lora的，后来想想，一直都是发布checkpoint的，那还是延续这个习惯吧。

我一直偏爱胶片摄影的光影色彩，因为那是无数美学工程师用百年的调色经验沉淀下来的美学的体现，是用图片记录光影色彩、记录美的最佳的、受众最广泛的体现。美不是被定义的，而是被感受的，能被定义的或者量化的，一定是会沦为平庸。所以美无法量化无法评估，只能去理解和感悟。美也无法让所有人都共情，能遇到一个人理解你，已经是万幸之事。

本次更新主要优化Z的人像类皮肤纹理和环境纹理细节，并优化模型生成人像类图像的美学。Z 本身生成的图片真实度非常高，某些情况下会让你完全分不清是AIGC图还是真实照片。但是，在美学方面，Z 还有优化提升的空间。

本次更新的模型整体风格略偏胶片人像摄影，可以生成很美的人像类图片，也可以生成很优美的环境细节，同时保留了 Z 的部分真实感。在提示词引导下可以生成非常真实的图片。

注意，因本次训练以合成的高质量人像摄影风格图片为主，在提升整体美学与纹理细节的同时，有可能会对复杂肢体动作和某些非人像类纹理造成影响。后续版本会持续优化。

欢迎体验试用，欢迎多多返图。

本次发布 BF16 & FB8 双版本，FP8 版8G显存可以使用，笔记本亲测。推荐使用 BF16 版，细节比FP8版丰富很多。

同步发表于魔搭：https://modelscope.cn/models/Nurburgring/BEYOND_REALITY_Z_IMAGE/summary

模型同步huggingface：
https://huggingface.co/Nurburgring/BEYOND_REALITY_Z_IMAGE

推荐参数：euler+simple，10~15步，CFG：2

BEYOND REALITY，I'm back。

圣诞&新年快乐！

最后，团队招新，我一手搭建起Z image数据生产团队，目前因数据需求不断增加，Z image数据生产团队开启招新，有意者加WX：Aptshtiso

欢迎加入我们Z image数据生产团队！

__________________________________________________________________________

Perhaps it is the best Z image model for portrait photography at present。

Long time no see. I miss you very much

The BEYOND REALITY series models have been out of update for more than half a year, and the mainstream open-source drawing models have also rapidly iterated to Generation Z.

Z IMAGE has brought everyone a lot of surprises, with its super-fast speed and ultra-realistic visuals. The calls for edit in the community have been continuous.

I know you're in a hurry, but don't worry yet. Let's take a look at the community's fine-tuned Z image first. Hahaha

BEYOND REALITY Z IMAGE 1.0 looks good with both heavy and light makeup. This is a model that is fine-tuned based on the portrait dataset and merged with lora. It continues the characteristics of the BEYOND REALITY series models, featuring high clarity and high aesthetics. One of the current problems of Z IMAGE is its poor performance of high-frequency information, which is reflected in the picture as poor texture details. Therefore, a lora was trained to improve the texture quality of the model. I originally intended to release lora, but then I thought that since I've always been releasing checkpoint, I might as well continue this habit.

I have always been fond of the light, shadow and color of film photography, because it is the embodiment of aesthetics accumulated by countless aesthetic engineers with a century of color-grading experience. It is the best and most widely accepted way to record light, shadow and color, as well as beauty through pictures. Beauty is not defined but felt. What can be defined or quantified is bound to become mediocre. So beauty cannot be quantified or evaluated; it can only be understood and perceived. Beauty cannot resonate with everyone. It is already a great fortune to meet someone who understands you.

This update mainly optimizes the skin texture and environmental texture details of Z's portrait type, and improves the aesthetics of the portrait type images generated by the model. The images generated by Z itself have a very high degree of authenticity. In some cases, you can't tell at all whether they are AIGC images or real photos. However, in terms of aesthetics, Z still has room for improvement and refinement.

The overall style of the model updated this time is slightly more like film portrait photography. It can generate very beautiful portrait pictures as well as very elegant environmental details, while retaining some of the realism of Z. Very realistic pictures can be generated under the guidance of prompt words.

Note that as this training mainly focuses on high-quality synthetic portrait photography style images, while enhancing the overall aesthetics and texture details, it may also have an impact on complex body movements and certain non-human portrait textures. Subsequent versions will continue to be optimized.

Welcome to try it out and send back more pictures.

This release features both BF16 and FB8 versions. The FP8 version can be used with 8GB of video memory. I have personally tested it on my laptop.

euler+simple，10~15步，CFG：2

Merry Christmas & Happy New Year!

BEYOND REALITY, I'm back.

_______________________________________________________________________________________________

解决了一些已知问题。

在这次升级过程中，多元化的艺术风格让我吃尽了苦头，折磨了我两个月。忽然发现，多姿多彩的艺术世界是多么的广袤精彩，有太多闻所未闻的艺术风格。本模型制作过程中，最初的尝试是为FLUX增加更多的、精彩的、多样化的艺术风格与提升模型审美。但是很明显，这次更新，我没能实现这个想法。虽然未达预期，也算是开启了艺术化之路，本模型系列将继续持续优化（希望下一个版本能达到预期）。在F1模型中，增加多样化的艺术风格会降低绘图的精度。。。可能是训练集不够？如果有个几百万图的多风格训练集洗模型，或许能在提升精度的同时，也提升多风格与审美？

本模型是并不是一系列的迭代过程中我最喜欢的模型，却是相对平衡的模型。能实现一定的艺术风格化创作，同时保持相对精致的细节。我最喜欢的是一个不稳定的模型，它经常犯错，却可以创作出出人意料的作品。所以，什么是错误，什么是正确？

本模型请尝试使用多种艺术风格进行引导，或许你也会有一些收获。即使这是一个折中的模型，也加入了一些艺术属性。

本模型适合画美女，嗯，因为女人最美。

2025.04.18

Fixed some known problems.

During this upgrade process, the diverse artistic styles caused me a lot of trouble and tortured me for two months. Suddenly, I realized how vast and wonderful the colorful art world is, with so many artistic styles that have never been heard of before. During the process of making this model, the initial attempt was to add more wonderful and diverse artistic styles to FLUX and enhance the aesthetic appeal of the model. But obviously, in this update, I failed to realize this idea. Although it did not meet expectations, it can be said that it has embarked on the path of artistry. This model series will continue to be optimized (hoping that the next version can meet expectations). In the F1 model, adding diverse artistic styles will reduce the accuracy of the drawing... Could it be that the training set is insufficient? If there is a multi-style training set washing model with several million images, perhaps it can improve both accuracy and multi-style and aesthetic appeal?

This model is not my favorite among a series of iterative processes, but a relatively balanced one. It can achieve a certain degree of artistic stylized creation while maintaining relatively exquisite details. My favorite is an unstable model. It often makes mistakes but can create unexpected works. So, what is wrong and what is right?

Please try to guide this model with multiple art styles. Maybe you will also gain something. Even though this is a compromise model, some artistic attributes have been added.

This model is suitable for drawing beautiful women. Well, because women are the most beautiful.

2025.04.18

贺蛇年，贺春节~

本次更新尝试提升图片的细节与质感。

修复了一些已知问题。

或许会有新问题，欢迎反馈。

祝各位朋友们中国春节快乐！

本想在除夕夜发布，一直上传失败。。。不知道C站出了什么问题。

希望你多多返图~

Congratulations on the Year of the snake and the Spring Festival.

This update attempts to improve the detail and texture of the image.

Fixed some known issues.

There may be new issues, please feedback.

I wish all friends a happy Chinese New Year!

I wanted to release it on New Year's Eve, but it failed to upload... I don't know what's wrong with civitai.

Hope you return more pictures ~

*****************************************************************************

新年新升级，提升细节，提升质感。

新年新气象，祝大家新年快乐~

New Year upgrade, improve the details, improve the texture.

Happy New Year to you all

图片1920x1080直出，希望绘制高清图的可以尝试。

1920x1080 straight out pictures , if you want to draw high definition pic, try it.

这是一个超级合并模型，合并了很多优秀的模型，在理解能力与表达能力上达到了一个令我满意的高度。在此分享出来，希望你能喜欢。

希望你多多返图。

This is a super merge model that combines many excellent models and reaches a satisfactory level in terms of comprehension and expression. I share it here and hope you enjoy it.

I hope you return more pictures.

补充一下，在Flux模型使用时，虽然官方提示prompt不要超过750字（貌似是？不知道有没有记错）但是在实际应用中发现，提示词写得越详细，对画面的描述越丰富，模型迭代出来的图像越精致。这表明，模型对超长提示词的理解非常好，响应也很好。我画图时尝试了3000字左右的提示词，出来的图很赞。如果你可以将画面描述得非常详细，那么，尽量就去详细的描述~

In addition, when using the Flux model, although the official prompt should not exceed 750 words (seems to be? However, in practical application, it is found that the more detailed the prompt words are written, the richer the description of the picture, and the more exquisite the image iterated out of the model. This shows that the model understands and responds well to very long prompt words. When I was drawing, I tried about 3000 words of prompt words, and the picture came out very good. If you can describe the picture in great detail, then try to describe it in great detail~

Files with SHA256: 5b5a67cb0d120a155172a6f03ff741a174ded387683bd3134911421d5eb8be6c