santa hat
deerdeer nosedeer glow
Sign In

One Trigger Word Lora Test: Warframe series-Garuda, Ivara, Equinox, Lotus, Citrine 一个触发词劳拉测试

0

Garuda - Warframe (nsfw)

Ivara - Warframe (nsfw)

Citrine - Warframe (nsfw)

Lotus - Warframe (nsfw)

Equinox (Merged form) - Warframe (nsfw)

Overall scores: Garuda 4.3(5), Ivara 4.3(5), Equinox 3.3(4), Lotus 3(3), Citrine 3.6(4)

Details: Garuda 5, Ivara 5, Equinox 3, Lotus 3, Citrine 5

All five LoRa have good details. The reversion of lotus and Equinox will be slightly more dependent on the model.

Style: 4

All five Loras can be switched freely from 2D to 3D. The 3D style is sometimes recognized as the skin and sometimes as the clothes. 2D effects rely on quality cues, and if the quality is high enough, it will be difficult to produce simple 2D images.

Failure rates: Garuda 4, Ivara 4, Equinox 3, Lotus 2, Citrine 2

Very interesting question, my conclusion is that there are traps in the trigger designed by the author.

The author uses the character name as a trigger word very intuitively, which can trigger character recognition well when relying on some dark or NFSW models. But there are many widely used models in my tests, and Lotus and Citrine inevitably lead to their common meaning. During my testing, I found that some parts of the character's body can also effectively guide the whole body information in LoRa, and I currently recommend using hemlet~

In addition, the 2D effect is not stable enough. This gave me an important tip, that is, a qualified model should be able to output any style, generate 2D when the details and quality requirements are not high, generate 3D when the requirements are high, and the most extreme details and quality will bring the same photo effect as taking photos.

总分:Garuda 4.3(5), Ivara 4.3(5), Equinox 3.3(4), Lotus 3(3), Citrine 3.6(4)

细节:Garuda 5, Ivara 5, Equinox 3, Lotus 3, Citrine 5

五个lora都有着不错的细节。lotus和Equinox的还原对模型的依赖会稍微大一些.

风格:4

五个Lora都能在2D到3D间自由切换。3D风格有时会被识别为皮肤,有时会被识别为衣服。2D效果依赖于质量提示词,如果质量够高会难以生成简单的2D图像。

故障率:Garuda 4, Ivara 4, Equinox 3, Lotus 2, Citrine 2

非常有意思的问题,我的结论是作者设计的trigger里有陷阱。

作者非常直观地使用角色名称作为触发词,在依赖一些黑暗或NFSW模型时可以很好触发人物识别。但是在我的测试里有很多广泛使用的模型,Lotus 和Citrine 不可避免地会被引导向他们的常见含义。我在测试时发现,一些人物身上的部位也能有效引导出lora中的全身信息,我目前最推荐使用的是hemlet~

另外,2D效果不够稳定。这给了我一个重要提示,那就是一个合格的模型应该能够输出任意画风,细节和质量要求不高的时候生成2D, 要求高的时候生成3D,最极致的细节和质量会带来拍照一样的照片效果。

Thanks to these Loras! Very interesting test !

0