JRU（The journey remains unfinished）道途未尽2.5

很高兴向大家介绍JRU（The journey remains unfinished）道途未尽2.5。该模型以Illustrious-xl为基底，重新训练的。本来是打算做JRU3的,结果一个负责抽象概念的分部炸炉了,将其整合进模型会导致生成的图片虽然每一个细节都正确,但是就是给人一种非常混乱的感觉。所以我直接去掉了那个部分,JRU2.5继承了JRU系列的一贯特点,即白色不止是白色。但是由于损失了一部分训练集,一些抽象的概念无法实现,(比如JRU2的封面图无法复刻)。尽管如此,我仍然认为这是一个优秀的,值得发布的版本,相比于JRU2,这个版本的肢体和稳定性有着跨越式的提高,同时在绘制横图大场景的时候,他能带来更精彩的构图,不止如此,JRU2所不擅长的NSFW内容以及多人构图,JRU2.5也可以轻松实现。由于使用了Illustrious-xl为基底,这么模型也会认识许多角色,使用画师tag也会产生奇妙的化反,大家可以用Danbooru的tag尝试一下。之后会尝试以这个模型作为基底训练画风.虽然c站对于横图的展示并不是很友好,但是我在创作例图的时候还是创作了大量的横图内容,因为这是JRU2.5相比于其他模型所擅长的,我希望大家可以好好欣赏.

另外在使用ADetailer对面部进行修复的时候,会导致面部同质化的问题.这不是模型的问题,不过在创作大场景时,还是推荐使用ADetailer

I'm excited to introduce to everyone JRU2.5 (The journey remains unfinished). This model is retrained based on the Illustrious-xl base. Initially, it was intended to be JRU3, but a division responsible for abstract concepts malfunctioned, causing the integration of this part to make the generated images correct in every detail yet appear very chaotic. Consequently, I removed that part. JRU2.5 retains the consistent characteristic of the JRU series, where white is more than just white. However, due to the loss of some of the training set, some abstract concepts cannot be realized (for example, the cover image of JRU2 cannot be replicated). Despite this, I still believe this is an excellent and worthwhile version to release. Compared to JRU2, this version has significantly improved limbs and stability, and it excels in composing horizontal, large scenes, bringing more exciting compositions. Furthermore, JRU2.5 can easily handle NSFW content and multi-person compositions that JRU2 struggled with. Since it uses Illustrious-xl as the base, this model also recognizes many characters, and using artist tags can produce fascinating effects. Everyone can try using Danbooru tags. I will attempt to train the painting style based on this model. Although civitai is not very friendly for displaying horizontal images, I still created a large number of horizontal contents when making example images, as this is what JRU2.5 excels at compared to other models, and I hope everyone can enjoy it.

Additionally, using ADetailer for facial repairs can cause issues with facial homogenization. This is not a problem with the model itself, but I still recommend using ADetailer when creating large scenes.

以下是第三次更新内容

Below is the content of the third update.

jru2的NSFW补丁已经上线，地址是https://civitai.com/models/349485
The NSFW patch for JRU2 is now live. The link is https://civitai.com/models/349485

在之后的发布作品中，我会发布许多用该模型训练的画风lora。这个模型画风单一，认识的角色也不多，但是它本身具有独特的画风，并且也很适合作为训练画风lora的base model。是的，我正是为此才弄出了这个模型。同时JRU2具有极其强大的，画背景的能力。
In my future releases, I will publish many style LoRAs trained using this model. This model has a singular style and recognizes few characters, but it has a unique style of its own and is also very suitable as a base model for training style LoRAs. Yes, this is exactly why I created this model.At the same time, JRU2 possesses an extremely powerful capability for drawing backgrounds.

About how to use this model

Ummmm, actually I am a very lazy person, so this user guide was written by a kind-hearted person who helped me out. I just copied it over.Let's thank him.

Regarding quality words

Quality words: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality,

Formula: Quality word + Character + Character appearance + Character action + Scene description + Other additional descriptions

This model's feature is that by changing negative prompt words, you can adjust the color palette of the image.

For example, you can directly apply a large number of negative prompt words from SD1.5, and the saturation of the image colors will be significantly increased, as shown below:

Negative: NSFW, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, (ugly:1.331), (duplicate:1.331), (morbid:1.21), (mutilated:1.21), (tranny:1.331), mutated hands, (poorly drawn hands:1.5), blurry, (bad anatomy:1.21), (bad proportions:1.331), extra limbs, (disfigured:1.331), (missing arms:1.331), (extra legs:1.331), (fused fingers:1.61051), (too many fingers:1.61051), (unclear eyes:1.331), lowers, bad hands, missing fingers, extra digit, bad hands, missing fingers, (((extra arms and legs)))

With the positive prompts remaining the same, using the simplified prompts of SDXL will make the overall image appear grayer, featuring a more realistic characteristic.

Negative: (low quality, worst quality:1.2), bad anatomy, fewer digits, extra fingers

With the positive prompts unchanged, not inputting any negative prompts can also result in normal image output, with the style as shown in the following

By adding 'chibi' to the positive prompts, you can then change the character's age through weights

Of course, you can also use as many wild and stunning words as possible to make the image richly artistic and gorgeous, just like I did with my cover.

If you want a more logical image, please use SDXL DPO-Turbo-LoRA (https://civitai.com/models/237775?modelVersionId=269354). If you wish for your image to be crazier and more gorgeous, please do not use it.

关于如何使用这个模型

emmmmmm其实我是一个很懒的人，所有这份使用说明是一个热心群众帮忙写的，我只是复制了过来，图片上面都放了 ，让我们感谢他。我觉得 没必要再放一次，大家就对照着看就行了哈。

质量词：masterpiece,best quality,great quality,good quality,normal quality,low quality,worst quality, 公式：质量词 + 角色+ 角色样貌+角色动作+场景描述+其他额外描述 这个模型的特点是通过改变负面提示词可以调整画面色彩， 例如，你可以直接套用大量sd1.5的负面提示词，画面色彩饱和度会得到大幅提升， 如下图： 负面： NSFW, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, (ugly:1.331), (duplicate:1.331), (morbid:1.21), (mutilated:1.21), (tranny:1.331), mutated hands, (poorly drawn hands:1.5), blurry, (bad anatomy:1.21), (bad proportions:1.331), extra limbs, (disfigured:1.331), (missing arms:1.331), (extra legs:1.331), (fused fingers:1.61051), (too many fingers:1.61051), (unclear eyes:1.331), lowers, bad hands, missing fingers, extra digit,bad hands, missing fingers, (((extra arms and legs))),

正面提示词不变，使用SDXL简化的提示词时画面整体会变灰，具有写实特点 负面： (low quality, worst quality:1.2),bad anatomy,fewer digits,extra fingrers,

正面提示词不变，不输入任何负面提示词也能正常出图

在正面提示词中加入【chibi】 ，然后可以通过权重来改变角色年龄

当然你也可以尽可能的使用一些疯狂且惊艳的词语来让画面变得富有艺术感的华丽，就像我封面做的那样

如果你想要更更有逻辑的画面，请使用SDXL DPO-Turbo-LoRA (https://civitai.com/models/237775?modelVersionId=269354),如果你希望你的画面更疯狂，更华丽，请不要使用他

以下是第二次更新内容

Below is the content of the second update.

我发现在xl模型的融合中，不同的模型只需要很少的比例带来的变化就是巨大的，所以我以0.1的比例在base model中融合了kohakuxl-delta-pre.safetensors。同时，我还重构了最后作为修正画风的lora，并且这一次没有再使用SDXL DPO-Turbo-LoRA（https://civitai.com/models/237775?modelVersionId=269354），这让我的模型更适合作为lora训练的base model。但是在生成图片是时候，我还是建议大家使用它。

I've found that in the fusion of XL models, different models only need a very small proportion to bring about significant changes, so I fused kohakuxl-delta-pre.safetensors into the base model at a ratio of 0.1. At the same time, I've also restructured the final LoRA used for style correction, and this time I did not use SDXL DPO-Turbo-LoRA (https://civitai.com/models/237775?modelVersionId=269354), making my model more suitable as a base model for LoRA training。However, when generating images, I still recommend using it.

例图大部分是直接使用该模型生成的图片，也有使用了我用该模型训练的lora生成的图片，直接使用该模型肢体可能相比与其他模型稍差一些，但是和使用该模型训练的画风lora一起使用之后，肢体会大幅度得到改善（我也不知道是什么原理，知道的可以告诉我为什么）

Most of the example images were generated directly using this model, and some were generated using the LoRA I trained with this model. Using this model directly, the limb quality might be slightly inferior compared to other models, but when used together with the style LoRA trained on this model, the limb quality significantly improves (I'm not sure of the principle behind this; if anyone knows, please tell me why).

感谢琥珀提供的kohakuxl-delta-pre.safetensors，在未发布时让我作为融合素材，推荐大家也去尝试她的新模型 Kohaku-XL Delta（https://civitai.com/models/332076）

Thanks to Kohaku for providing the kohakuxl-delta-pre.safetensors, which allowed me to use it merger before its release. I recommend everyone to try her new model, Kohaku-XL Delta （https://civitai.com/models/332076).

以下是第一次更新的内容

Below is the content of the first update

为什么要尝试融合XL底模？因为高山仰止，景行行止，虽不能至，心向往之。

Why attempt to merge XL models? Because we hold the highest peaks in awe and cherish the journey's path; though we may not arrive, our hearts aspire to it.

我需要一个只有一种画风的xl模型，这样我就可以用它来训练不同画风的ora附加在它的上面，让他能够更稳定的通过lora表现出更多的画风，因此我便开始了尝试。

I need an XL model with a singular style so I can train it with various styles of ORA added on top, allowing it to more stably express a wider range of styles through LoRA. Thus, I began my attempts.

它的画风完美的符合我个人的审美，并且无论是NSFW内容还是SFW内容，它都可以胜任

Its style perfectly matches my personal aesthetic, and it is capable of handling both NSFW and SFW content.

它目前尚不完善，人物的肢体崩坏率要比其他的xl模型更高一些，我会尝试继续完善它。除此之外，我认为这是一个非常棒的模型。

It is still imperfect, with a higher rate of limb distortion in characters compared to other XL models. I will attempt to continue refining it. Aside from this, I believe it is an excellent model.

接下来是融合方法

The next step is the integration method.

将Animagine XL V3（https://civitai.com/models/260267?modelVersionId=293564）与Kohaku-XL gamma（https://civitai.com/models/270291?modelVersionId=317831）用MBW以以下比例融合，得到Mix2。

To blend Animagine XL V3（https://civitai.com/models/260267?modelVersionId=293564） with Kohaku-XL （https://civitai.com/models/270291?modelVersionId=317831）gamma using MBW at the following ratio, resulting in Mix2.

之后再将得到的Mix2与Aurora XL base(https://civitai.com/models/301957/aurora-xl-basepony-xl)进行融合，配方如下

Afterward, the obtained Mix2 is blended with Aurora XL base(https://civitai.com/models/301957/aurora-xl-basepony-xl), following the recipe below.

最后一步是训练，使用了三个500张图片的训练集，每个训练集的数据分别有着不同的画风特征，将他们分别训练成lora以后，整合进模型中，此外还在模型中整合了SDXL DPO-Turbo-LoRA (https://civitai.com/models/237775?modelVersionId=269354)

The final step is training, using three training sets each with 500 images, where each set has distinct artistic style characteristics. After training them into LoRA separately, they were integrated into the model. Additionally, SDXL DPO-Turbo-LoRA (https://civitai.com/models/237775?modelVersionId=269354) was also integrated into the model.