If you like what you see, feel free to leave a comment, send pictures, and more in the comments section below!

请看每个版本的模型简介！

• 【猫娘乐园】Nekopara-AllGirl LORA v1 animagineXLV3epoch-2

  是一款基于stable diffusion XLLORA技术训练的模型，旨在创造出高质量的Nekopara系列角色图像。我在A40显卡上使用7000张精选Nekopara图片经过10小时的训练，虽然遇到了轻微的过拟合和欠拟合问题，但通过适当调整权重，最终达到了令人满意的效果。基础模型使用了animagineXLV3_v30.safetensors模型，保证了图像的清晰度和细节表现。

• 我推荐的最低运行分辨率为768*768，以确保图像的细节不会丢失。在训练过程中，我采用了全标签训练方法，以实现更精准的角色控制和图像生成。为了更稳定地生成特定角色，我设计了一套触发提示词系统。主要提示词为【nekopara】，辅助以角色特定的提示词，例如【Vanilla_(nekopara),Nekopara_Vanilla】等，配合详细的场景描述，可以生成极具魔法氛围的高质量角色图像。

• 此外，我还推出了v2版本，特化于hans系列模型，基于hansv27推土机v1.1版本训练，针对hans系列的SDXL模型进行了优化，特别是在双人图像生成方面取得了显著进展。通过增加到8400张的数据集，v2版本在处理双人场景时更为稳定，支持使用多个角色提示词来精准控制角色的出现。同时，我也修复了v1版本中的欠拟合问题，并通过数据集的扩展，优化了图像的多样性和质量。

• 【猫娘乐园】Nekopara-AllGirlLORA系列模型，无论是单一角色的精细刻画，还是复杂场景下的角色互动，都能提供高质量的图像输出，为用户带来极致的视觉享受。无论是作为创作工具，还是用于个人娱乐，本系列模型都是Nekopara粉丝和AI艺术爱好者的理想选择。

Please see the model introductions for each version!

• Nekopara-AllGirl LORA v1 animagineXLV3epoch-2

  This model, based on the stable diffusion XL LORA technology, was trained to create high-quality images of characters from the Nekopara series. It was trained using 7000 selected Nekopara pictures on an A40 graphics card over 10 hours. Despite encountering minor overfitting and underfitting issues, satisfactory results were achieved through appropriate weight adjustments. The base model utilized the animagineXLV3_v30.safetensors model, ensuring the clarity and detail of the images.

• The minimum recommended resolution for running this model is 768x768 to ensure no loss of detail in the images. Throughout the training process, a full-label training method was adopted for more accurate character control and image generation. To more reliably generate specific characters, I designed a trigger keyword system. The main keyword is "nekopara", supported by character-specific keywords, such as "Vanilla_(nekopara), Nekopara_Vanilla", along with detailed scene descriptions, to generate high-quality character images with a magical atmosphere.

• Moreover, I also introduced a version 2, specialized for the hans series models, trained based on the hansv27 bulldozer v1.1 version. This version was optimized for the hans series' SDXL models, especially making significant progress in generating images of two characters. With an expanded dataset to 8400 images, version 2 is more stable in handling scenes with two characters, supporting the use of multiple character keywords for precise control of character appearances. Additionally, I fixed the underfitting issues found in version 1 and enhanced the diversity and quality of images through dataset expansion.

• The Nekopara-AllGirl LORA series models, whether for the detailed portrayal of individual characters or the interaction of characters in complex scenes, provide high-quality image outputs, offering users an ultimate visual experience. These models are an ideal choice for Nekopara fans and AI art enthusiasts, whether used as a creative tool or for personal entertainment.

计划等数据集攒到1w左右的高质量nekopara图片后再进行大更新