WJMJ——Sword Armor&Wings——万剑魔甲（神兵玄奇）

2024.12.10 update:

The first "usable" Lora model after Pony training (SD1.5 failed after multiple attempts).

Currently, only the front non-special actions have no obvious defects(but the hairstyle cannot be changed).

Due to the quality of the material, ADetailer must be used, otherwise the face will be distorted.

If you want to change the color, please add a color before "armor", but not all colors are guaranteed to be available.

Because the material is too limited and the training rules are not clear, it is currently impossible to make further improvements.

————————————————————————————————————————

5.6 update:

Use the revsoap28D_v11 large model to retrain a version, and the material itself has not been changed.

The effect: a little more realistic, but not by much.

Since normal results still cannot be obtained in other large models, it is still defined as: unusable

————————————————————————————————————————

An experimental training about AI painting, in order to explore the basic algorithm of AI painting.

If you are interested in the prototype, you can search for the Chinese name.

Based on the perfectWorld_v2Baked large model training, because I tried SD1.5, ChilloutMix, Anything V5 as the basic model, the result was even worse.

The base resolution is 512*1024.

Without any experience, I can only try a little bit according to the tutorial. The current conclusion is that AI cannot recognize things that are completely absent in its own database. Therefore, the learning process of AI painting cannot even conduct a detailed analysis of the training objects.

At present, it is completely impossible to see that AI painting has any deconstruction algorithm for human body or three-dimensional space.

For me, this model is almost unusable,because it is too limited,such as it can only be used for the basic resolution, the pose is single, even with the pose model, it is almost impossible to see, there is no model versatility, and so on.

Just ask someone to do better,if it arouses the interest of professionals,I hope to make a usable model, I would be very grateful.

2024.12.10更新：

改用Pony训练后第一个“可用”的Lora模型（SD1.5尝试多次均失败）。

目前只有正面非特殊动作无明显缺陷（但发型无法变更）。

由于素材质量问题，必须使用ADetailer，否则脸部会崩坏。

如想改变颜色，请在“armor”前加色彩，但不保证所有色彩可用。

因为素材过于有限，且不清楚训练规律，目前无法作进一步完善。

————————————————————————————————————————

5.6更新：

用revsoap28D_v11大模型重新训练一版，素材本身未作改动。

效果是：更写实了一点，但不多。

由于在其它大模型中依旧不能出正常结果，故依然定义为：不能用。

————————————————————————————————————————

关于AI绘画的一个实验性质的训练，为了探索AI绘画的基础算法。

如果对原型感兴趣，可以搜索中文名。

基于perfectWorld_v2Baked大模型训练，因为尝试过SD1.5、ChilloutMix、Anything V5做基础模型，结果更糟。

基础分辨率为512*1024。

没有任何经验，只能根据教程一点一点试，目前得出的结论是：AI是认不清自己数据库内完全没有的事物的，故AI绘画的学习过程甚至都做不到对训练对象的完全拆解。

目前也完全看不出AI绘画对人体，或者对立体空间有任何解构的算法。

此模型对我来说，几乎是不可用的，因为局限太大，比如只能用于基础分辨率，姿势单一，甚至加上openpose也几乎没法看，没有模型泛用性，等等。

只求抛砖引玉，如果引起了大神的兴趣，希望能做出可用的模型，不胜感激。