<h1 id="-t2k2mjwgi">请仔细阅读描述</h1><h3 id="please-read-the-description-carefully.-x31kr73ip">Please read the description carefully.</h3>Scroll down to see the description in English.我非常喜欢<a target="_blank" rel="ugc" href="https://civarchive.com/models/639170/virile-stallion">Virile Stallion</a>这个模型。它在生成真实的男性角色图像时拥有无与伦比的效果，无论是SFW还是NSFW。不过，我注意到，这个模型似乎不能很好地理解和生成具有过长包皮的阴茎的图像，所以我训练了这个lora。这是我第一次尝试训练lora，所以这个lora模型存在许多问题和局限，如果你在使用中觉得有什么不对劲，欢迎在评论区提出意见，我也许会在之后尝试改进，但是介于我繁忙的学业，这可能需要相当长的时间🥲<h3 id="lora-i67d74vx9">以下是一些使用建议，我强烈建议你按照建议使用此lora。</h3><ol><li>这个lora以 Virile Stallion V5.0 Photoreal （<a target="_blank" rel="ugc" href="https://civarchive.com/models/639170/virile-stallion">https://civarchive.com/models/639170/virile-stallion</a>） 作为底模进行训练，所以理论上它在配合Virile Stallion V5.0 Photoreal 使用时效果最佳。你也许可以尝试与其他真实风格的Pony模型一起使用，但可能不会获得很好的图像</li><li>建议模型强度为0.7。更高可能会显著劣化生成的画面</li><li>建议在正向提示词中添加“uncut”、“foreskin”</li><li>训练所用的数据集均为真实图像，所以你最好不要将其用于动漫风格的pony模型上。而且，动漫和二次元风格的Pony模型通常都可以较好地生成包皮过长的阴茎，不需要再额外使用lora。</li><li><ul><li>对于1.0版本的模型，训练所用的图像分辨率多为1024*1536。所以请务必将生成图像的原始大小设定为1024*1536或更高，之后再进行高清修复。若是将图像大小的分辨率设定为低于1024*1536并使用此lora，将会产生很差的图像。我知道这个分辨率对于大多数人的电脑来说有点过高了，但由于我是第一次训练lora，所以不知道会发生这种问题🥲。我在2.0版本的模型中略微降低了数据集的分辨率，如果你的电脑配置不甚理想，可以试试2.0版本。</li></ul><ul><li>对于2.0版本的模型，训练所用的图像分辨率多为960*1280。所以请将生成图像的原始大小设定为960*1280或更高，之后再进行高清修复。768*1024的分辨率也可以尝试，但生成的图像质量不太好。</li><li>为什么不将数据集图像分辨率降低到1024*1024或768*1024？我觉得包皮是个很复杂的结构，1024*1024分辨率的图像对于阴茎的特写照片也许拥有足够的细节，但是数据集中应该需要包含一些人物的全身照片，否则Lora会使生成的图像趋近于特写。对于全身照片，较低分辨率的图像会让阴茎的包皮细节丢失，我觉得这或许并不利于lora生成好的图像......当然，我对lora模型训练的技术细节一窍不通，若你觉得这样的看法有失偏颇，请随意指正。</li></ul></li><li>推荐的设置：<ul><li>图像分辨率：1024*1536（对于1.0版本的模型）/960*1280（对于2.0版本的模型）</li></ul><ul><li>采样器：DPM++</li><li>降噪（对于ComfyUI）：0.9</li><li>CFG强度：5.5</li><li>采样步数：32</li><li>模型强度：0.7</li><li>高清修复/放大 倍数：1.05或更高</li><li>高清修复/放大 采样步数：29或更高</li></ul>这些设置应该允许在最低的性能要求下生成较好的图像。如果你希望得到更高分辨率的图像，大可随意拉高放大倍数和采样步数。</li><li>使用Adetailer（对于A111 WebUI）或FaceDetailer（对于ComfyUI）对画面进行局部重绘在大多数情况下都能改善阴茎的细节。<ul><li>对于A111 WebUI的用户，你可以使用 Adetailer 插件和 <a target="_blank" rel="ugc" href="https://civarchive.com/models/132388?modelVersionId=149639">Penis + Pussy Adetailer Model</a> 来检测和重绘</li><li>对于ComfyUI的用户，你可以使用 FaceDetailer 节点和 <a target="_blank" rel="ugc" href="https://civarchive.com/models/301092/cock-and-ball-detection-adetailer">Cock and Ball Detection </a>模型来检测和重绘。在以下链接中，模型的作者给出了详细的使用说明：<a target="_blank" rel="ugc" href="https://civarchive.com/models/301092/cock-and-ball-detection-adetailer">https://civarchive.com/models/301092/cock-and-ball-detection-adetailer</a></li></ul></li><li>训练所用的图像部分含有背景和完整的人脸，所以这个Lora会对生成图像的人物面部和背景产生一定的影响。降低模型强度或是使用提示词固定人物面部特征和背景特征可能会有所改善。在所有版本中都容易出现光头的男人（这并不是说我的训练集中的某个光头的男人不好看🥲），你可以尝试在负面提示词中加入“bald”缓解此现象。</li><li>这个lora应该能同时处理勃起和疲软状态下包皮过长的阴茎的形态。你可以使用“erection”（勃起）或“flaccid”（松弛）提示词来控制阴茎的形态。</li></ol>最后，希望这个模型能对你有所帮助。<hr />I don't speak English very well. I hope you don't mind 🥲I really like the <a target="_blank" rel="ugc" href="https://civarchive.com/models/639170/virile-stallion">Virile Stallion </a>model, this model has unparalleled results in generating realistic images of male characters, both SFW and NSFW.However, I noticed that this model doesn't seem to be able to understand and generate images of circumcised penis very well, so I trained this lora. this is my first attempt at training a lora, so there are many problems and limitations with this model, if you find anything wrong in using it, feel free to leave a comment and I may try to improve it later, due to my busy school schedule, this might take quite a long time 🥲<h3 id="here-are-some-suggestions-that-i-highly-recommend-you-follow-with-this-lora.-soybfdlsd">Here are some suggestions that I highly recommend you follow with this lora.</h3><ol><li>This lora is trained with Virile Stallion V5.0 Photoreal （<a target="_blank" rel="ugc" href="https://civarchive.com/models/639170/virile-stallion">https://civarchive.com/models/639170/virile-stallion</a>）as the base model, so theoretically it works best when used with Virile Stallion V5.0 Photoreal. You may try to use it with other realistic style Pony models, but you may not get good images.</li><li>a model strength of 0.7 is recommended. higher strengths may significantly degrade the resulting image.</li><li>Recommend adding “uncut”, “foreskin” to the positive prompts.</li><li>the datasets used for training are all real images, so you'd better not use them on anime style Checkpoints. Also, anime style Pony models usually do a better job of generating circumcised penises without the need for additional lora.</li><li><ul><li>The image resolution used for training is mostly 1024*1536. so please make sure to set the original size of the generated image to 1024*1536 or higher, and then do Hires fix or upscale. If you set the resolution of the image size to less than 1024*1536 and use this lora, it will produce a very poor image. I know that this resolution is a bit too high for most people's computers, but since this is the first time I've trained the lora, I didn't know this problem would occur 🥲. I've slightly reduced the resolution of the dataset in version 2.0 of the model, so if your computer's performance is less than stellar, try version 2.0.</li></ul><ul><li>For version 2.0 of the model, the image resolution used for training is mostly 960*1280. so please set the original size of the generated image to 960*1280 or higher, and then do Hires fix or upscale. 768*1024 resolution can also be tried, but the quality of the generated image is not so good.</li></ul><ul><li>Why don't I lower the dataset image resolution to 1024*1024 or 768*1024? I think the foreskin is a very complex structure, and a 1024*1024 resolution image might have enough detail for a close-up photo of the penis, but the dataset should need to include some full-body photos of the people, or else Lora will make the generated images tend to be close-ups. For the full body photos, the lower resolution images will make the foreskin detail of the penis lost, which I think might not be conducive for lora to generate good images ...... However, I don't know anything about the technical details of lora model training, so please feel free to correct me if you feel this is a biased view.</li></ul></li><li>Recommended settings:<ul><li>Image resolution: 1024*1536 (for v1.0 models) / 960*1280 (for v2.0 models)</li><li>Sampler: DPM++</li><li>Denoise (for ComfyUI): 0.9</li><li>CFG : 5.5</li><li>Sampling steps: 32</li><li>Model Strength: 0.7</li><li>Hires fix/upscale: 1.05x or higher</li><li>Hires fix/upscale Sampling Steps: 29 or higher</li></ul>These settings should allow better images to be produced with minimal performance requirements. If you want a higher resolution image, feel free to increase the magnification and sampling steps.</li><li>Localized inpainting of images using Adetailer (for A111 WebUI) or FaceDetailer (for ComfyUI) improves penis detail in most cases.<ul><li>For A111 WebUI users, you can use the Adetailer plugin and <a target="_blank" rel="ugc" href="https://civarchive.com/models/132388?modelVersionId=149639">Penis + Pussy Adetailer Model</a></li><li>For ComfyUI users, you can use the FaceDetailer node and the <a target="_blank" rel="ugc" href="https://civarchive.com/models/301092/cock-and-ball-detection-adetailer">Cock and Ball Detection Model</a>. In the following link, the author of the model gives detailed instructions on how to use it: <a target="_blank" rel="ugc" href="https://civarchive.com/models/301092/cock-and-ball-detection-adetailer">https://civarchive.com/models/301092/cock-and-ball-detection-adetailer</a></li></ul></li><li>The image used for training partially contains background and full faces, so this lora will have some influence on the faces of the people and the background in the generated images. Reducing the model strength or using prompts for faces and backgrounds may help. Bald men tend to appear in all versions (which is not to say that one of the bald men in my training set isn't a good look 🥲), and you can try to mitigate this by adding “bald” to the negative prompts.</li><li>The lora should be able to handle both erect and flaccid uncircumcised penises. You can use “erection” or “flaccid” prompts to control the shape of the penis.</li></ol>I hope this model will be helpful to you.

Uncut For VirileStallion

此版本略微降低了数据集中图像的分辨率。现在你应该可以用更低分辨率的图像大小来生成比1.0模型更好的结果，比如960*1280。更高分辨率的图像依然是更好的选择，1024*1536的分辨率也可以使用，并且通常会得到更好的结果。This version slightly reduces the resolution of the images in the dataset. You should now be able to generate better results than the 1.0 model with lower resolution image sizes, such as 960*1280.Higher resolution images are still a better choice, 1024*1536 can also be used and will usually give better results

Files with SHA256: 2b12e958776fce172de52059bfecde8b28068fdc3e93d5043776a3877d66484d