DreaMirror-Mix 梦镜 – v2.0

中文/English

这是一个画风更偏向于2D的2.5D混合模型，如果有需要改进和做得不对的地方请指出来（如果你能说出纠正的方法最好，因为我可能不是很清楚）。

这个模型的方向是尽可能忠于提示词,能生成更多的东西(人物是主要的,但不代表其他事物的效果不重要),舒服的光影。

你可以用它尝试任何东西！我会不定时补充预览图。如果你觉得这个模型不错,可以评价一下以及上传你的图片。

从AOM3:https://huggingface.co/WarriorMama777/OrangeMixs 和Anything:https://civitai.com/models/9409/or-anything-v5 知道了很多模型和学习了很多混合方面的经验,十分感谢。

使用建议:

如果在脸部不是焦点的画面中(非正面也是同样),脸部出现崩坏的情况,以及想提升面部质量,使用局部重绘 重绘区域使用!!仅蒙版!!(非常建议此方法)获得更好的面部,或使用Hires. fix改善,使用其他随机种或者工具也是不错的办法。(我认为这是混合了3D或者2.5D模型导致的,并没有什么很好的办法解决,如果你有办法在模型层面改善请告诉我。)

较高的分辨率(比512 * 512高一点)再加上Hires. fix,图片质量会更好(如果显存不够你可以尝试低倍率的Hires. fix或者其他放大方法)。

过高的,用于画面质量(像best quality这些)的正面和负面提示词是不必要的,会减少细节,也可能使画面趋于一种风格,以及无法对提示词做出反应。

如果觉得画面内容不够丰富,你可以尝试细致地描述,使画面更加贴近你的想象。

如果想要更强的明暗对比效果,可以尝试光影类lora,比如epi_noiseoffset:https://civitai.com/models/13941/epinoiseoffset (我曾在V1加入过它,但是它似乎会影响手指生成,所以去掉了,你可以自己使用它)

如果你想要更多的亚洲面孔,人物类lora是很好的选择,比如Doll Likeness系列:https://huggingface.co/Kanbara/doll-likeness-series

This is a style more 2D preference for 2.5D merge model, so please point out if there is something that needs to be improved and done wrong (If you can say how to correct the best, because I may not be very clear).

The direction of this model is to be as faithful as possible to the prompt words, to be able to generate more things (the character is the main, but does not mean that the effect of other things is not important), comfortable light and shadow.

You can try anything with it! I will add preview images from time to time. If you think this model is good, comment on it and upload your pictures.

I know a lot of models and learn a lot of mixing experience from AOM3:https://huggingface.co/WarriorMama777/OrangeMixs and Anything:https://civitai.com/models/9409/or-anything-v5, thank you very much

Suggestions for use:

If the face is not in focus in the Image (same for non-frontal), the face appears to be falling apart, and you want to improve the quality of the face, use Inpaint and Inpaint area use!!only Masked!! (This method is highly recommended) to get a better face, or use Hires. fix to improve, use other seed or tools is also a good way.(I think this is because of the merge of 3D or 2.5D models,There is no good way to fix it,If you have a way to improve it at the model level please let me know.)

Higher resolution (a little higher than 512 * 512) plus Hires. fix, picture quality will be better (if the gpu memory is not enough you can try a Low magnification of Hires. fix or other upscale tools)

Overly high, positive and negative prompt for image quality (like best quality) are unnecessary and will reduce detail, and may tend to make the image look a specific style and the inability to respond to the prompt.

If you feel that the content of the picture is not rich enough, You can try to describe in detail to make the picture more closely to your imagination.

If you want a stronger contrast between light and dark, you can try light and shadow lora, such as epi_noiseoffset:https://civitai.com/models/13941/epinoiseoffset (I had added it in V1, but it seems to affect the finger generation, so remove it, you can use it yourself)

If you want more Asian faces, the Character class lora is a good choice, such as the Doll Likeness series: https://huggingface.co/Kanbara/doll-likeness-series

你可以从版本信息中知道当前版本信息,比如使用了哪些模型/You can know the current version information from the version information, such as which models are used

我使用这个VAE/I use this VAE:https://huggingface.co/stabilityai/sd-vae-ft-mse-original

配方在/Recipe in → huggingface:https://huggingface.co/Fre2C/DreaMirror-Mix

查看更多