模型出图结果对比

• Mi-Ha-V2：较大程度地提升了出图的逻辑性与泛化性，在画风上最接近原画。

• ACertainThing：效果一般。

• Anything：效果较好，中规中矩。

• momoko-p：有着momoko特有的服饰褶皱以及阴影，画面更加的炫彩。

• OrangeMix：出图失败（放大后依然很糊），不过目测潜力巨大。

使用说明

• 禁止商用，如果在其他插画网站发布该模型生成的图片，请标记AI，我非常反感不打标记的行为。

• 大部分说明和Himitsu - HaneRu | Stable Diffusion Checkpoint | Civitai一致。

• 放大算法：R-ESRGAN 4x+ Anime6B 默认参数即可。

• Unet权重：

  • Mi-Ha-V2为 0.3 左右

  • 其余四个模型为 1 左右

  • 未经测试的模型请自行调参

• TEnc权重均为0较好。

• 由于训练集的缺陷，男人的手会出现在人物旁边，请使用负面prompt屏蔽，具体prompt参见各版本内容。

Comparison of model results

• Mi-Ha-V2: A greater degree of logic and generalization out of the picture, the closest to the original in terms of drawing style.

• ACertainThing: The effect is average.

• Anything: better effect, moderate.

• momoko-p: with momoko unique clothing folds and shadows, the picture is more colorful.

• OrangeMix: the picture failed (still very paste after enlargement), but the visual potential is huge.

Istructions

• Commercial prohibited, If you post images generated by this model on other illustration sites, please mark AI, I am very much against unmarked behavior.

• Most of the instructions are the same as Himitsu - HaneRu | Stable Diffusion Checkpoint | Civitai.

• Amplification algorithm: R-ESRGAN 4x+ Anime6B Default parameters are sufficient.

• Unet weights:

  • Mi-Ha-V2 is about 0.3

  • The remaining four models are around 1

  • Please refer to the untested models by yourself

• TEnc weights are all 0 is better.

• Due to the defect of the training set, the man's hand will appear next to the character, please use negative prompts to mask, See the contents of each version for specific prompts.

离自产自销又近了一步（确信）