Sign In

SanaeXL-anime

17
85
2
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
85
Reviews
Published
Jul 17, 2024
Base Model
SDXL 1.0
Trigger Words
masterpiece,best quality,highres
Hash
AutoV2
EC08BC2B85

中文:
这是一个基于Kohaku XL EPS rev3改进的图像生成模型。我们使用了约300万张图像进行训练,主要改善了原版在多数场景下的脚部生成问题,并尝试加入了一些新的艺术风格。虽然还有进步空间,但希望这些改进能为用户带来更好的使用体验。

使用方法:与Kohaku EPS rev3相同,建议按以下顺序输入提示词:1girl + 人物名 + 作品名 + 风格 + 其他描述正面提示词请加入: masterpiece, bestquality, highres
负面提示词必须包含: lowres (由于使用了大量低分辨率图片进行训练)
worst quality 和 low quality 可根据个人需求选择是否添加本模型仍有诸多不足之处。恳请使用者提供宝贵反馈,我们将继续努力改进。

特别鸣谢kohaku训练的基础模型与deepghs开源的数据集。

English:
This is an image generation model improved upon Kohaku XL EPS rev3. We trained it with approximately 3 million images, focusing on enhancing foot generation in various scenarios and introducing some new artistic styles. While there's still room for improvement, we hope these enhancements will provide users with a better experience.

Usage: Same as Kohaku EPS rev3. Suggested prompt order: 1girl + character name + work title + style + other descriptionsPositive prompts: Please add masterpiece, bestquality, highres
Negative prompts: Must include lowres (due to training with many low-resolution images)
worst quality and low quality can be added based on personal preferenceThis model still has many areas for improvement. We kindly request user feedback to help us continue enhancing it.

Special thanks to the base model trained by kohaku and the open-source dataset from deepghs.

日本語:
これはKohaku XL EPS rev3をベースに改良した画像生成モデルです。約300万枚の画像を使用して学習し、主に元のバージョンで見られた多くのシーンでの足の生成問題を改善し、いくつかの新しい芸術スタイルを取り入れました。まだ改善の余地はありますが、これらの改良がユーザーの皆様により良い体験をもたらすことを願っています。

使用方法: Kohaku EPS rev3と同様です。プロンプトは次の順序で入力することをお勧めします: 1girl + キャラクター名 + 作品名 + スタイル + その他の説明ポジティブプロンプト: masterpiece, bestquality, highres を追加してください
ネガティブプロンプト: lowres を必ず含めてください (多くの低解像度画像を使用して学習したため)
worst quality と low quality は必要に応じて追加してくださいこのモデルにはまだ多くの改善点があります。ユーザーの皆様からのフィードバックをお待ちしております。引き続き改良を重ねてまいります。

kohakuが学習した基本モデルとdeepghsのオープンソースデータセットに特別な感謝を申し上げます。

Discussion

GromForever's Avatar
GromForever

I was going to mention that the text in the description was hard to read because it was semi-transparent, but I see that's already been fixed. Also, please update the tags for the model. The current tags don't seem to help at all. they don't describe the model accurately. About the model itself, I don't even know what to say. The basic Kohaku Epsilon can produce good images when it comes to SFW content, but NSFW results are still terrible. Overall, my opinion is that if you need an image of a specific character in a variety of styles, this model works great. However, if you try to add a second or third character to the scene or make it more complex, the results become very unstable, and you often end up with mutants. It's possible that I might have been doing something wrong, maybe the tags were in the wrong order, or my requests to the model were too complex. However, I have other models to compare it to, and the results here are disappointing.