本次的模型是符华的三合一lora，主要包括了风语鸥鸣（泳装）、太虚山仙忆和月轮。

风语鸥鸣的prompt为：

Fu Hua(Seagull),1girl,swimsuit,sunglasses,ponytail, single glove,

太虚山仙忆的prompt为：

Fu Hua(Taixu),1girl,chinese clothes,black gloves,

其中“black gloves”对应的是手臂上的黑色绷带，使用该词有会加大手臂（手指）紊乱的概率，如对细节要求不高可以不使用该词。

月轮的prompt为：

Fu Hua(Knight),1girl,bodysuit,pongtail,

其中将“outfit,”加入negative prompt中可以减少画面中出现奇怪的手脚（装甲）的出现概率。

为了减少出图的3D质感你可以尝试把“3d rendering,”加入到negative prompt里。

该lora推荐的权重为0.7-0.9，但调用月轮模型时需将其调高至0.8-1.0。

此模型基于“animefull-final-pruned.ckpt”，不同大模型适配程度见图8。

针对脸不像符华的问题，在本次的训练中我加入了不少脸部的特写，你可以试着在生成图片的过程中可以把“face,”也加入到关键词中去。但生成大头照时有概率会出现月轮的头饰，为此你需要将有关月轮的关键词放入negative prompt中。

该模型泛用性不好，想换装的可以试试最开始那个云墨的，那个泛用性其实还行。

以下是无关的碎碎念：

虽然我之前就听说lora尽量不要做多合一的，但我就很想试试。为此其实上一个青鸟模型一开始就是二合一的，但由于两套服装过于相似，并且考虑到泛用性的问题，最终我还是将其拆分成了两个。

而这个模型一开始我其实打算做五合一的，但最终因为种种原因只成为了三合一，其中最大的原因便是训练时间太长了，之前一版模型一般只要30-60min，而这次则直接到了4h，并且在3.5h的时候电脑直接宕机了（悲）。再加上后期的整理验证等原因，光一版就至少要花掉我一整天的时间，所以之后多半不会再出多合一的模型了。

感谢一直看我碎碎念到这里的你！最后预告一下，下一个模型应该是符华的“玄衣素裳”。

This version of the model is a three-in-one compilation featuring Fu Hua in different outfits: "Seagull's Soar" (swimsuit), " Taixuan Impression" ,and " Shadow Knight" .

For the " Seagull's Soar" prompt:

Fu Hua(Seagull),1girl,swimsuit,sunglasses,ponytail, single glove,

For the "Taixuan Impression" prompt:

Fu Hua(Taixu),1girl,chinese clothes,black gloves,

(Please note that the mention of "black gloves" may slightly increase the chance of irregularities in the appearance of the arms or fingers. If precise details are not of utmost importance, the mention of gloves can be excluded.)

For the "Shadow Knight" prompt:

Fu Hua(Knight),1girl,bodysuit,pongtail,

By adding the phrase "outfit" to the negative prompt, it is possible to minimize the occurrence of peculiarities in the depiction of her limbs or armor.

 Additionally, including "3D rendering" in the negative prompt can help reduce the three-dimensional effect in the generated image.

Recommended weight:0.7-0.9

(When using the " Shadow Knight " should increase to 0.8-1.0)

This model is based on the "animefull-final-pruned.ckpt",and the adaptation degree of different models can be seen in Figure 8.

To address the issue of Fu Hua's face not resembling her character, the training process included numerous close-up shots of her face. You may try including the prompt "face" during the image generation process. However, when generating headshots, there is a possibility of the inclusion of Shadow Knight's headpiece. To avoid this, prompt related to Shadow Knight should be added to the negative prompt.

It is worth mentioning that this model's versatility is limited. If you are interested in outfit changes, the previously mentioned "Azure Empyrea" model may be more suitable as it offers better versatility.

On an unrelated note:

Although it has been advised against creating multi-character models with Lora, I couldn't resist giving it a try. In fact, the previous "Cerulean Court" model was initially intended to be a combination of two characters but was eventually split into two due to the similarity of their outfits and concerns regarding versatility.

As for this model, I originally intended to include five variations but ended up with only three due to various reasons. The main reason was the excessively long training time. While previous versions typically took between 30 and 60 minutes, this one took 4 hours. Unfortunately, my computer crashed at the 3.5-hour mark, resulting in much longer processing time. Considering the subsequent organizing and verifying tasks, creating just one version required me to devote an entire day. Consequently, it is unlikely that multi- outfit models will be developed in the future.

Thank you for patiently reading my nagging so far! As a sneak peek, the next model will feature Fu Hua in the " Rustic Noir " outfit.

English translated by ChatGPT.