Ann-Pony-3D-LCM
详情
下载文件 (2)
关于此版本
模型描述
The following is a translation by Youdao, my English is very poor → _ →：
Set -A1111 or Forge
Dragging the example image into the picture information transfer diagram is the fastest(forge can not read the Dynamic Thresholding information of the picture, must manually set it on the other side (~ ▽ ~ "),
Manual setting:
1. Select the model
2. Select SDXL-VAE
3. Sampler LCM, scheduler: automatic or Karras or Polyexponential
4.CFG: 7-8(1.5~2 if you don't like to use Dynamic Thresholding)
4. Adjust the resolution: the resolution has a great impact on XL and pony, sometimes the map is not good to change the resolution)
704x1408 (0.5)|704x1344 (0.52)|768x1344 (0.57)|768x1280 (0.6)|832x1216 (0.68)|832x1152 (0.72)|896x1152(0.78) My favorite is ||896x1088 (0.82)|960x1088 (0.88)|960x1024 (0.94)|1024x1024 (1.0)|1024x960 (1.07)|1088x960 (1.13)|1088x896 (1.21)|1152x896 (1.29)|1152x832 (1.38)|1216x832 (1.46)|1280x768 (1.67)|1344x768 (1.75)|1344x704 (1.91)|1408x704 (2.0)|1472x704 (2.09)|1536x640 (2.4)|1600x640 (2.5)|1664x576 (2.89)
|1728x576 (3.0)
5. Open the Dynamic Thresholding, Mimic CFG Scale: 5~6, Mimic CFG Scale and CFG Scale choose Half Cosine Up, the lower limit of the function choose 1.5~2 (I do not understand the principle of these parameters, are based on the figure to see the effect to judge roughly up or down a little bit will affect what, and then find their own think the appropriate value), Perhaps many people do not love to use it, its general role should be to make the prompt words more obedient, the picture is more rich
6. Because LCM will cause the screen to turn gray. So turning on BMAB to adjust contrast and color is the only way to do it, unless you incorporate a LORA that can increase image saturation without affecting other aspects
BMAB(Just found that forge is not supported, forge will collapse after installation, turn over the table! (╯ □ ')╯︵┻ - ┻ forge):
portu-sim/sd-webui-bmab: Auto masking and inpainting for person, face, hand. Resizing image using detection model. (github.com)
设置-A1111或者Forge
拖例图进图片信息转文生图最快(forge无法读取图片的Dynamic Thresholding信息，必须手动在那边设置一下╮(￣▽￣")╭ )，手动的话：
1.选择模型
2.选择SDXL-VAE
3.采样器LCM,调度器：自动或者Karras或者Polyexponential
4.CFG: 7-8(如果不喜欢用Dynamic Thresholding就是1.5~2)
4.调整分辨率:（分辨率对XL及pony都影响很大，有时图不好换个分辨率即可）
704x1408 (0.5)|704x1344 (0.52)|768x1344 (0.57)|768x1280 (0.6)|832x1216 (0.68)|832x1152 (0.72)|896x1152(0.78)我最喜欢这个||896x1088 (0.82)|960x1088 (0.88)|960x1024 (0.94)|1024x1024 (1.0)|1024x960 (1.07)|1088x960 (1.13)|1088x896 (1.21)|1152x896 (1.29)|1152x832 (1.38)|1216x832 (1.46)|1280x768 (1.67)|1344x768 (1.75)|1344x704 (1.91)|1408x704 (2.0)|1472x704 (2.09)|1536x640 (2.4)|1600x640 (2.5)|1664x576 (2.89)
|1728x576 (3.0)
5.开启Dynamic Thresholding，Mimic CFG Scale：5~6，Mimic CFG Scale和CFG Scale都选择Half Cosine Up，函数下限选择 1.5~2 （这些参数原理我不懂，都是凭出图看效果来判断大致调高或调小一点会影响什么，然后找到自己认为的合适值），也许很多人不爱用它，它的大致作用应该是可以让提示词更听话，画面更加丰富
6.因为LCM会导致画面变灰。。所以开启BMAB调高下对比度和颜色是唯一的办法了，除非融合一个可以提高画面饱和度而不影响别的方面的LORA
BMAB(刚发现forge他喵的又不支持，forge一装出图就崩，翻桌！(╯‵□′)╯︵┻━┻ forge)
portu-sim/sd-webui-bmab: Auto masking and inpainting for person, face, hand. Resizing image using detection model. (github.com)
The following is a translation by Youdao, my English is very poor → _ →：
Hello everyone, I am a helpless pain designer from CH, INTP personality o(' • ~ • '). o Because of the illegal behavior in CH, I declare that the release of this model is solely for learning and communicating AI image generation technology, and shall not be used for various illegal laws and regulations of various countries! I have been lurking in Station C for almost a year, and I have blended some models before, but I did not publish them because I did not think there was anything special. However, ever since I discovered the Pony model, I have fallen in love with it because it is really free and performs well! So, I made up a model that fits my aesthetic. ツ
2.5D_Prony - v1.5 | Stable Diffusion Checkpoint | Civitai I started with the ppkkmoon model, because I prefer the 3D game texture of the picture - the main way is to integrate the three LORAs (by layering)
2.photo 【pony】 - photo_v1 | Stable Diffusion LoRA | Civitai
3.Detail Tweaker XL - v1.0 | Stable Diffusion LoRA | Civitai
Thank you for your sharing!
Specific Settings, download my picture, and then load the picture information, self-feel fine tuning, is the best way. Because my WebUI Settings used a lot of parameters: ADetailer +Dynamic Thresholding (CFG Scale Fix)+BMAB+LoRA Block Weight (LoRA Block Weight)+ dynamic prompt words
Of course, my graph parameters are my past year of accumulated graph semen, poof wrong, experience, you must also have their own habits and skills, so welcome back to graph exchange ~
Oh, by the way, the Pony model is the same as the XL model, the hardware requirements for the graph in the WebUI are very high, and my 4080 graphics card can only be amplified 1.5 times and often crash =. After Ming, I will share my comfyUI process, which has relatively low hardware requirements and is more playable
Finally, it is necessary to emphasize that domestic players should not casually carry the model of C station, and release it to those websites that make money by model in China, which is to lose the spirit of open source.
大家好，我是来自CH的一名苦逼设计师，INTP人格o(` • ~ • ′。)o 因为在CH搞色色犯法,所以我声明发布此模型完全是用于学习和交流AI图像生成技术，不得用于各种违法各国法律法规的事情！我在C站潜伏快一年了，之前也融过一些模型，但没发布，因为我觉得没有什么特别之处。但是，自从我发现了Pony模型，我便爱上了它，因为它真的很自由，表现也很棒！所以，我就自己调配了一个符合我审美的模型~o(*≧▽≦)ツ
2.5D_Prony - v1.5 | Stable Diffusion Checkpoint | Civitai 我从ppkkmoon大佬这个模型开始，因为我更喜欢3D游戏质感的画面~主要方式是融入了这3个LORA（通过分层的方式）
2.photo 【pony】 - photo_v1 | Stable Diffusion LoRA | Civitai
3.Detail Tweaker XL - v1.0 | Stable Diffusion LoRA | Civitai
感谢这些大佬们的分享！
具体设置的话，下载我的图片，然后加载图片信息，进行自我感觉的微调，是最好的方式了。因为我的WebUI设置参数用了很多:细节修复(ADetailer)+Dynamic Thresholding (CFG Scale Fix)+BMAB+LoRA 权重分层设置(LoRA Block Weight)+动态提示词
当然，我的生图参数都是我过去一年所积攒的生图精液，噗不对，经验而来，各位大佬一定也有自己的习惯和技巧，所以欢迎返图交流~
哦，对了，这个Pony模型和XL模型一样，在WebUI里生图对硬件要求很高，我4080显卡也只能是放大1.5倍还经常崩=。=，所以各位量力而行╮(￣▽￣")╭ 我之后会分享我的comfyUI流程，那个对硬件要求相对较低，而且可玩性更强~
最后，要着重强调国内玩家不要随便搬运C站的模型，发布到国内那些靠模型赚钱的网站，那是失去开源精神的令人唾弃的！