Sign In


Checkpoint Merge
Jun 23, 2023
Base Model
SD 1.5

  • feel free to buy me a coffee ☕

  • XXMixunreal:我的2.5D新模型,一个极具特色的模型,可以根据关键词进行变种,创造属于自己的风格化图片。A highly distinctive model that can generate variations based on keywords, creating personalized, stylized images.

  • XXMix_Petrichor:这是我新的系列模型,比较偏网红图片质感一些,希望大家可以支持下载尝试一下

  • XXMix_Petrichor:This is my new series of models, which are more focused on the aesthetic of internet celebrity images. I hope everyone can support and download them to give them a try


You can try this high-definition restoration algorithm. I believe that in some pictures, the effect of using this magnification algorithm is better, and the picture is more stylized. The link to the high-frequency (hf) is below, and those who are interested can give it a try.






我认为修手最好的是Detail Tweaker LoRA (细节调整LoRA),其他都是垃圾0 0


  • 图片信息里的XXMix_4_v2893模型就是v30版本,图片是在模型改名前生成的

  • 1.很高兴向大家介绍XXMix_9系列的3.0版本。这个版本在泛用性和对LORA的支持上都有明显的加强,并且手部的表现要明显好于之前的版本。3.0版本是由v2.5、v2.6和XXMix_4这三个版本Mix而来。生成的图片我尽量没有用任何Embedding模型,因为我发现有些Embedding模型比较隐晦,你不知道哪些Embedding模型起了作用哪些没起作用,从而不能还原图例。

  • 2.如果你想还原图例,请先安装以下扩展:

  • Adetailer:建议在生成图片时始终开启(如果你想90%以上还原图例,那么建议你在图片信息内查看adetailer是否已开启)。

  • 注意:Adetailer需要下载一些模型,这里是模型地址。模型存放路径:webui/models/adetailer。没有找到adetailer文件夹自己建一个即可。

  • Adetailer视频教程,这里我就贴娜乌斯嘉的教程喽

  • CFG Scale Fix:有需求的时候开启,会提高画面的质感。我用的默认数值7左右,如果是白天CFG数值可以稍微大一些,晚上数值要小一些。

  • (经过我的测试,在第二台电脑上没有办法100%还原图例,我现在还不知道是什么原因,有可能是CFG Scale Fix这个扩展的问题,也有可能是分块VAE的问题,不过基本上你可以还原到90%-95%左右,对画面的影响不是很大。)

  • 3.对于生成图片,个人的小心得:如果你得到了一张非常喜欢的图片,但是手部或其他部位的问题很严重,如果你不想用局部重绘修改的话,那么建议你在固定好seed之后,微调CFG的数值(每次调整0.2-0.3这种幅度来尝试),这样你就有可能在保持原图的基础上得到一张完美的图片。

  • 4.在这里,我还想分享一个我喜欢的后缀句子,它可以让画面的表现力更强。当然,你可以根据需要修改句子中的单词来微调画面风格。这些单词包括:fantasy, high contrast, ink strokes, explosions, over exposure, purple and red tone impression , abstract, ((watercolor painting by John Berkey and Jeremy Mann )) brush strokes, negative space,

  • 5.最后,如果你喜欢这次模型的更新,希望你可以把生成的好看图片分享出来。

Version 3.0 Description:

  • The XXMix_4_v2893 model in the image information is actually the v3.0 version, but the images were generated before the model was renamed.

  • Pleased to introduce the 3.0 version of the XXMix_9 series. This version has significantly improved in versatility and support for LORA, and the hand performance is significantly better than the previous versions. The 3.0 version is a mix of v2.5, v2.6, and XXMix_4. When generating images, I try not to use any Embedding models as much as possible because I have found that some Embedding models are more obscure, and you don't know which ones are working and which ones are not, which makes it difficult to restore the example.

  • If you want to restore the example, please install the following extensions first:

  • Adetailer: It is recommended to always enable it when generating images. If you want to restore the example by more than 90%, it is recommended to check whether Adetailer is enabled in the image information.

  • Note: Adetailer requires downloading some models from Here is the model address. The model should be stored in the path: webui/models/adetailer. If you cannot find the adetailer folder, please create one yourself.

  • Adetailer Chinese Video Tutorial,这里我就贴娜乌斯嘉的教程喽

  • CFG Scale Fix: Enable it when needed, it will improve the quality of the image. I use the default value of about 7. If it is daytime, the CFG value can be slightly larger, and at night, the value should be smaller. (After my test, I cannot restore the example 100% on the second computer. I still don't know the reason. It may be a problem with the CFG Scale Fix extension or the Tiled VAE. However, you can basically restore it to about 90%-95%, and the impact on the image is not significant.)

  • For generating images, here is my personal experience: If you get a picture that you really like, but there are serious problems with the hands or other parts, and you don't want to use local redraw to modify it, then it is recommended to adjust the CFG value slightly (0.2-0.3 range) after fixing the seed. This way, you may get a perfect picture while keeping the original image.

  • Here, I would like to share a suffix sentence that I like, which can make the image more expressive. Of course, you can modify the words in the sentence to fine-tune the image style. These words include: fantasy, high contrast, ink strokes, explosions, overexposure, purple and red tone impression, abstract, brush strokes ((watercolor painting by John Berkey and Jeremy Mann)), negative space.

  • Finally, if you like the updates of this model, we hope you can share the beautiful images you generated.


v26-fp16-no-ema update




Version 2.6 release notes:

This version is a minor update that optimizes the face and hand issues. However, the main focus is still on the gacha system. The major issues will be addressed in the next version.




V2.5 version has better lighting effects and supports more fantasy content. The basic face shape has been adjusted, and the performance in scenes has been improved.


适当调小(CFG Scale)的数值,可以得到与众不同的效果。





4.如果想得到更好的皮肤质感,那么生成图片的时候不要开启高清修复,而是将生成的图片传入图生图模式,启用Tiled Diffusion,来进行放大,从而得到更真实的皮肤质感。



Version 2 Release Notes:

  1. Changed the default face.

  2. Improved lighting and shadow effects.

  3. Added more models and will continue to iterate in the future.

  4. For better skin texture, do not enable Hires Fix when generating images. Instead, use the "Tiled Diffusion" mode to enlarge the generated image and achieve a more realistic skin texture.

  5. I will continue to update and iterate on this large model, hoping to add more content and make it more interesting.

Thank you for your support!



Recommended Parameters:

Sampler: DPM++ 2M Karras alt Karras or DPM++ SDE Karras

Steps: 20~40

Hires upscaler:4x-UltraSharp

Hires upscale: 2

Hires steps: 15

Denoising strength: 0.2~0.5

CFG scale: 6-8

clip skip 2