🌍欢迎加入QQ群"兔狲的SDXL梦工厂",群号 :835297318(入群答案:兔狲)。新建的Telegram群聊“兔狲的SDXL百老汇”,链接:https://t.me/+KkflmfLTAdwzMzI1
🤝如您有模型定制或其他合作需求,请联系Wechat: 13439222668。请简单注明来意,时间宝贵不闲聊哈,谢谢理解!
👩🎨Unleash the power of Xmirror.ai, the ultimate tool for generating ultra-realistic human images. With unlimited generations, bring your creative vision to life and only use credits when you download satisfactory images. Use my promo code: LEOSAM30 to get a 30% discount at www.xmirror.ai
📖这是用于生成胶片风格AI写真照片的LoHA & LoRA模型。与majicMIX realistic、MoonMIX或Chilloumix搭配使用,可以生成逼真的胶片风格照片。同时推荐搭配TEXTUAL INVERSION模型:bad_pictures、negative_hand Negative Embedding。这两个Negative Embedding模型不会显著影响到胶片的质感。
This is LoHA & LoRA models for generating film style AI portrait photos. when used with majicMIX realistic、 MoonMIX or Chilloumix, it can generate realistic film style photos. At the same time, it is recommended to use the TEXTUAL INVERSION model: bad_pictures, negative_hand Negative Embedding. These two Negative Embedding models will not significantly affect the film texture.
更新内容:
Velvia 3.0的主体LoRA模型的训练集图像数量已达到406张,分辨率进一步提升至1024*1024。
本版本主体模型训练使用了Scale weight norms功能,以避免模型过拟合。在另一个备用版本中我也测试了“Prodigy” Optimizer, 这个Optimizer可以在一半的训练总步数下提供与AdamW8bit相当的拟合程度,但可能是我参数设置不合适,训练结果的人脸美感较差,因此最终没有采用。
在高分辨率主体模型基础上,少比例融合了Velvia 2.0模型,同时进行了分层调节,以改善人像肤色。
Update Content:
The number of training images for the main LoRA model of Velvia 3.0 has reached 406, with the resolution further increased to 1024x1024.
In this version, the main model training utilizes "Scale weight norms" functionality to avoid overfitting. I also tested the "Prodigy" Optimizer in another alternative version. Prodigy can provide a fitting degree comparable to AdamW8bit with half the total training steps, but perhaps due to inappropriate parameter settings, the facial aesthetics of the training results were not as good, so it was not adopted in the end.
Based on the high-resolution main model, a small proportion of the Velvia 2.0 model was blended, and hierarchical adjustments were made to improve the skin tone in portraits.
更新内容:
Velvia 2.0的主体LoRA模型的训练集图像分辨率进一步提升至768*1152。
本版本主体模型训练时,加入了15张亚洲年轻男性胶片训练照片,以提高模型在生成男性照片时的真实感。
在高分辨率主体模型基础上,少比例融合了之前两版评价良好的Filmgirl历史版本模型。经过多次测试,这样融合模型的操作可以减少LoRA模型对脸型的影响。
Updates include:
The training set image resolution of the main LoRA model in Velvia 2.0 has been further improved to 768x1152.
During the training of this version's main model, 15 film training photos of young Asian males were added to enhance the realism of generated male photos.
On the basis of the high-resolution main model, a small proportion of the well-reviewed historical versions of Filmgirl were incorporated. After multiple tests, this combination of models can reduce the impact of the LoRA model on facial contours.
VELVIA模型是采用高分辨率训练集以及新的打tag方法训练得到的LoRA模型。
更新内容:
照片训练集进行了重新筛选,不再使用正则化图像
训练集分辨率由512x704提升至640x960
采用了新的打tag方法与金字塔噪声
The VELVIA model is a LoRA model trained using a high-resolution dataset and a new tagging method.
Updates include:
The photo training dataset has been reselected, no longer using regularized images.
The training dataset's resolution has been increased from 512x704 to 640x960.
A new tagging method and pyramid noise has been adopted.
我常用的CFG范围是6~7,Lora权重通常在0.5~0.7。
Clip skip设为1或2时,生成的效果会有所不同,大家可以多尝试看自己喜欢哪种风格。
Please note:
My commonly used CFG range is 6 to 7, and the Lora weight is usually 0.5 to 0.7.
When the Clip skip is set to 1 or 2, the generated effects will be different. You can try more to see which style you like.
为免大家在下载使用胶片风LoRA和LoHA模型时产生混淆,未来这个模型的主页将专用于更新胶片风的LoRA模型。
To avoid confusion when downloading and using the film-style LoRA and LoHA models, the homepage for this model will be dedicated to updating the film-style LoRA model in the future.