Medusa (Rider) | Fate/stay night: Unlimited Blade Works

Medusa,<lora:chara_FateStayNightUBW_Medusa(Rider)_v1_1:1>,

medusa_ubw,

medusa_ubw,<lora:chara_FateStayNightUBW_Medusa(Rider)_v1_1:1>,

Total score: 2.3 (3 stars) available

Details: 2 Partially available

Some 2D models already have a small amount of detail about the character. Without using trigger words, or even without this lora, some facial and hair details can be triggered. When using trigger alone, 3D models tend to generate mythical snake-haired banshees, and 2D models are closer to characters in FSN. Since no other prompt was used, it can be seen that LoRa has successfully trained some of the clothing details, but the performance is unstable in different models. The main source of interference is the snake-haired banshee in the model. These detailed questions hint at the principles of writing tags and triggers, and the need of control sets.

Style: 3 Available

2D is too strong, but choosing the right 3D model can see some 3D style. Performing the worst in RealisticVision is somewhat surprising.

Failure rate: 2 Partially available

When producing the 2D style, there are many images of multiple human heads, as well as some text information. It is not clear whether the author chose the appropriate training set. Some images have unwanted weapons , which seems to be related to different trigger words? There are many problems, but it is not fatal for drawing the correct character.

The test used only trigger word, generating 40 images with 4 seeds and 10 models. Examine the initial performance of the character LoRa itself. There are three objective dimensions: detail, style, and failure. 1 Not available 2 Partially available 3 Available 4 Excellent 5 Perfect . The overall score matches Civitai's five-star rating.

总分：2.3（3星 ）可用

细节：2 部分可用

部分2D模型中已有该人物少量细节。不使用提示词，甚至不用该lora也能触发一些面部和头发的细节。单独使用trigger时，3D模型更倾向于生成神话中的蛇发女妖，2D模型和FSN中的角色更接近。由于没有使用任何其他提示词，可以看出lora中训练成功了一部分服饰细节，但是在不同模型中表现不稳定。主要干扰来源于模型中的蛇发女妖。这些细节的问题提示了tag和trigger的编写原则，以及对照集的作用。

风格：3 可用

2D过强，但选择合适的3D模型能看到一些3D风格。在realisticvision中表现最差有些令人意外。

失败率：2 部分可用

产生2D风格时出现了很多多个人头的图像，以及部分文本信息。不太清楚作者是否选取了合适的训练集。一些图像中生成了不需要的武器，似乎和trigger word有关？问题较多，但对于画出正确的角色不致命。

该测试仅使用trigger word，用4个种子和10个模型生成了40个图像。考察人物LoRa 自身的初始性能.有三个客观维度：细节、风格和失败。1 不可用 2 部分可用 3 可用 4 出色 5 完美 。总分和Civitai的五星匹配。