Type | |
Stats | 5,849 |
Reviews | (164) |
Published | Mar 3, 2023 |
Base Model | |
Training | Steps: 42,560 Epochs: 20 |
Trigger Words | elaina 颜色_style_costume 服装类型_costume witch_hat brooch boots broom wand |
Hash | AutoV2 99CBC59F7F |
宣传视频已推出，欢迎大家支持。
Video introduction has been launched, welcome to support.
视频地址：https://www.bilibili.com/video/BV16Y4y1y7h4
V3.0 中文介绍
请允许我以全新的的Lora发布elaina_v3，因为这将是一次完全的重制，并不是简单的升级，我是以制作最完美的伊蕾娜Lora为目的制作了它，这个Lora倾注了我大量的心血，即使不算租借GPU的费用，它也消耗了我整整7天的时间，关于制作的详情我将放到结尾，以下便是本Lora的详细使用说明。
使用说明：
依然写在最前面的，本Lora意在还原小说画风，使用它你并不能很好的生成动画画风与其他画风，经过我的调研，许多人更喜欢的是动画里的屑魔女，但很抱歉，我暂时没有去做，也许以后会尝试。
elaina_v3我最想先列出的新功能，那就是伊蕾娜的衣服风格，在elaina_v2中，只有魔女服和毛衣两种款式，而现在你可以根据季节与背景的色调，完全定制伊蕾娜的魔女服的款式或是日常服，我们有多种完全不同的风格可供选择：
基本的风格选择通式是：颜色_style_costume+服装类型_costume
颜色=blue、purple、black(成功率低)、pink、darkblue、cyan、crimson、magenta(成功率高)
服装类型=majo、daily
没错，你可以通过两个tag组合获得8种不同的魔女服风格，再加上不会改变风格的日常服，总共有9种风格。细心的你们应该能发现了，本Lora的预览图，9张图分别对应了9种风格，选择你们喜欢的风格吧。
定制了衣服以后伊蕾娜身上的其他组件(魔女帽witch_hat、扫帚broom、靴子boots)的风格不统一怎么办？没问题，elaina_v3中只要你输入了衣服风格通式：颜色_style_costume，再搭配任意伊蕾娜身上的组件，或是两个甚至三个的组合，它们都会尽可能地适配相应的魔女服风格。当然，你也可以给日常服的伊蕾娜配上组件，虽然日常服的风格不会改变，但相应的组件你依然可以自定义风格。
接下来便是最激动人心的新功能，我尽力还原了伊蕾娜的魔女胸针，它不再是elaina_v2里那样飘忽不定，时而完全崩坏，时而忽大忽小，使用brooch，你便可以近乎原版的生成它；如果你认为效果依然不满意，与之前一样，调大参数进行局部重绘，我保证你能生成一个完美的魔女胸针。同样，伊蕾娜的魔女帽与靴子也修复了崩坏的问题，你将很难再看见像开瓢的扇子一样崩坏的帽子和后跟不成形状的鞋子了，使用witch_hat、boots放心的生成它们吧。
我甚至对伊蕾娜的魔法杖也进行了还原，你可以通过wand或是holding_wand生成它。
最后一个功能，将是我这会投入最多时间的，那就是扫帚的修复，我甚至自己画了10几幅扫帚来补充素材。因为扫帚断裂成两半一直是个千古难题，最根本的原因就是AI学习的素材里，扫帚不是在身后挡着就是拿在手里，给AI造成了扫帚是两段的假象，这在每个大模型中也都存在这样的问题。为了在Lora中修复它，我需要给它学习大量“完整的”扫帚，而且素材里所有关于伊蕾娜以外的的所有区域我都要完全扣干净，防止被误认为是扫帚，但是最终的成果还是要靠运气，至少比上个版本好太多了，这也是唯一一个没能完美解决的问题，以后技术成熟后我会继续努力修复。如果你不想看见扫帚，请在负面提示词中输入broom、holding_broom、riding_broom，我保证你几乎不会再看见扫帚；同理要生成扫帚，请在正面提示词中输入它们。
制作详情：底模为Counterfeit-V2.5，训练总步数共42,560步，主要是部件的训练也需要很多步数。素材图片有94张，其中还有我自己画的图，小说画风的训练素材太少仍然是个难以逾越的鸿沟。
写在最后：感谢你能看到这里，为了制作这个Lora，我付出了太多，从一开始逐一修改tag，再到逐一抠图选材，都耗费了我大量的时间，我在抠图的时候看见满满文件夹的图片，我曾无数次的想要放弃，但elaina_v2中大家的支持让我坚持制作了下去，保证每一幅图扣的完美，每一个tag都无懈可击，每一个部件都做到各种角度与风格，最终我完成了它。当然最根本的原因还是我自己的技术力不够强，素材的质量不够高，但这也是无可奈何的，因为我的本职工作与AI技术毫无瓜葛，一切都是凭着自己的兴趣制作的，如果这个Lora能帮到你，给我一个5颗星评价吧，谢谢你们。
如果我的Lora取得成功，我将会上传一个b站视频进行宣传，到时候请你们多多支持了。
V3.0 English introduction
Please allow me to release elaina_v3 with the new Lora, because this will be a complete remake, not a simple upgrade. I made it for the purpose of making the most perfect Elaina Lora. This Lora has invested a lot of my efforts. Even if the cost of renting the GPU is not included, it also consumed my entire 7 days. I will put the details of the production to the end. The following is the detailed instructions for the use of this Lora.
Instructions for use:
It is still written at the top. This Lora intends to restore the style of the novel. Using it, you cannot generate animation and other styles of painting. After my research, many people prefer the Elaina in the animation, but I am sorry , I won't do it for now, maybe I will try it later.
The new function I want to list first in elaina_v3 is Elaina's clothing style. In elaina_v2, there are only two styles of witch clothes and sweaters. Now you can completely customize the style of Elaina's majo costume or daily costume according to the season and the color of the background. We have a variety of completely different styles to choose from:
The basic style selection formula is:颜色_style_costume+服装类型_costume(”颜色“ meaning color, "服装类型" meaning clothing type)
颜色=blue,purple,black(low success rate),pink,darkblue,cyan,crimson,magenta(high success rate)
That's right, you can get 8 different witch costume styles through the combination of two tags, plus the daily costume that don't change the style, there are 9 styles in total. Careful you should be able to find that the preview of this Lora, 9 pictures correspond to 9 styles, go and choose your favorite style.
What if the styles of other components (witch hat, broom brush, boots boots) on Elaina's body are not uniform after customizing the clothes? No problem, as long as you enter the clothing style selection formula in elaina_v3: 颜色_style_costume, and then match any components on Elaina, or a combination of two or even three, they will match the corresponding witch costume style as much as possible. Of course, you can also add components to Elaina in the daily costume. Although the style of the daily costume will not change, you can still customize the style of the corresponding components.
Next is the most exciting new function. I tried to restore Elaina's majo brooch. It is no longer elaina_v2, it is so erratic, sometimes completely collapsed, sometimes big or small. Using brooch, you can generate it close to the original; if you think the effect is still not satisfactory, as before, increase the parameters and use Inpainting, I guarantee you can generate a perfect witch brooch. Similarly, Elaina's hat and boots have also fixed the problem of collapse. It will be difficult for you to see hats that collapse like a fan and shoes with shapeless heels. Using witch_hat or boots to generate them safely.
I even reworked Elaina's magic wand, you can spawn it with wand or holding_wand.
The last function, which I will spend the most time on, is the restoration of the broom. I even drew 10 or so brooms myself to supplement the material. Because breaking the broom into two segments has always been an eternal problem. The most fundamental reason is that in the material of AI learning, the broom is either behind or in the hand, which creates the illusion that the broom is two segments for AI, which also exists in every Checkpoint. In order to repair it in Lora, I need to make it learn a lot of "complete" brooms, and I have to completely clean all the areas in the material except for Elaina, so as not to be mistaken for brooms, but the final result It still depends on luck, at least it is much better than the previous version. This is the only problem that has not been solved perfectly. I will continue to work hard to fix it after the technology matures. If you don't want to see brooms, enter broom, holding_broom, riding_broom as negative prompts and I guarantee you will almost never see brooms again; similarly to generate brooms, enter them as positive prompts.
Production details: The base model is Counterfeit-V2.5, and the total number of training steps is 42,560, mainly the training of components also needs a lot of steps. There are 94 material pictures, including pictures drawn by myself. There are too few training materials for the novel style of painting, which is still an insurmountable gap.
Write at the end: Thank you for seeing here. In order to make this Lora, It took me a lot of time to modify the tag one by one at the beginning, and then to sketch and select the material one by one. I saw the pictures full of folders when I was sketching. I wanted to give up countless times, but with the support of everyone in elaina_v2, I insisted on making it to ensure that every picture button is perfect, every tag is impeccable, and every component is made from various angles and styles. Finally, I finished it. Of course, the most fundamental reason is that my own technical ability is not strong enough and the quality of the material is not high enough, but this is also helpless, because my own work has nothing to do with AI technology. Everything is made based on my own interests. If this Lora can help you, give me a five-star rating, thank you.
If my Lora is successful, I will upload a video to promote it on Bilibili. When the time comes, please support me a lot.